Paper deep dive
Perception-Aware Communication Middleware for Distributed Visual Perception in UAV Swarms
Manveen Kaur, Kevin Loi, Ifunanya Okafor, Daniel Ng, Joseph Lucey-Renteria
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/22/2026, 4:29:00 AM
Summary
This paper presents a perception-aware communication middleware designed for distributed visual perception in UAV swarms. Unlike conventional packet-level QoS, this middleware treats complete perception data samples (images) as the unit of communication, addressing the need for low-latency, high-throughput transport of high-bandwidth data in resource-constrained, heterogeneous UAV environments. The system extends a lightweight UDP broker-based publish-subscribe architecture with services including image fragmentation/reconstruction, concurrent packet transmission via thread pooling, priority-aware scheduling with aging mechanisms, and middleware-level Image Quality Assessment (IQA) using the Variance of the Laplacian. Evaluated on a heterogeneous hardware testbed (Raspberry Pi, Intel NUC, NVIDIA Jetson AGX Orin) using YOLOv8n for object detection, the middleware demonstrates superior throughput, low end-to-end latency, effective traffic prioritization under load, and mitigation of detection degradation by filtering blurred images.
Entities (12)
Relation Signals (13)
Perception-Aware Communication Middleware → testedonhardware → NVIDIA Jetson AGX Orin
confidence 98% · A Raspberry Pi 4 (RPi4), Intel Next Unit of Computing (NUC), and NVIDIA Jetson AGX Orin (Orin) serve as U S , U L , and U Y v8 , respectively.
Perception-Aware Communication Middleware → testedonhardware → Intel NUC
confidence 98% · Intel Next Unit of Computing (NUC)... serve as U L
Perception-Aware Communication Middleware → testedonhardware → Raspberry Pi 4
confidence 98% · A Raspberry Pi 4 (RPi4)... serve as U S
Perception-Aware Communication Middleware → evaluatedwith → YOLOv8n
confidence 97% · The middleware is evaluated on a heterogeneous hardware testbed emulating a UAV swarm using YOLOv8n object detection.
Perception-Aware Communication Middleware → includesservice → Image Quality Assessment (IQA)
confidence 96% · The middleware extends... with perception-specific services, including... image quality assessment.
Perception-Aware Communication Middleware → includesservice → Image Fragmentation
confidence 96% · The middleware extends... with perception-specific services, including image fragmentation and reconstruction
Perception-Aware Communication Middleware → includesservice → Priority-Aware Scheduling
confidence 96% · The middleware extends... with perception-specific services, including... priority-aware scheduling
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Unmanned Aerial Vehicle (UAV) swarms increasingly support safety-critical applications that rely on distributed visual perception. Meeting the low-latency requirements of these applications can require perception models to execute within the swarm on inference-capable UAVs, creating a need for efficient UAV-to-UAV transport of high-bandwidth perception data. However, the Quality-of-Service (QoS) requirements of perception differ from conventional packet-level QoS; successful delivery of individual packets does not ensure that a complete, timely, and usable image is available for inference. We present a novel perception-aware communication middleware that treats complete perception-data samples as the communication objects for which QoS must be satisfied. The middleware extends a lightweight UDP broker-based publish-subscribe architecture with perception-specific services, including image fragmentation and reconstruction, concurrent packet transmission, priority-aware scheduling, and image quality assessment. The middleware is evaluated on a heterogeneous hardware testbed emulating a UAV swarm using YOLOv8n object detection. Experimental results demonstrate low end-to-end application latency, substantially higher throughput than a lightweight UDP broker, effective prioritization of perception traffic under increasing background load, and mitigation of object-detection degradation through middleware-level image quality assessment. This work provides an initial framework for integrating AI-specific data handling into communication middleware to support emerging distributed AI applications in multi-agent mobile cyber-physical systems.
Tags
Links
- Source: https://arxiv.org/abs/2609.24964v1
- Canonical: https://arxiv.org/abs/2609.24964v1
Trouble viewing inline? Open PDF directly →
Full Text
35,790 characters extracted from source content.
Expand or collapse full text
Perception-Aware Communication Middleware for Distributed Visual Perception in UAV Swarms Manveen Kaur ∗ Kevin Loi ∗ , Ifunanya Okafor ∗ , Daniel Ng ∗ , Joseph Lucey-Renteria † ∗ Department of Computer Science, California State University, Los Angeles Los Angeles, CA, USA mkaur39, kloi2, iokafor6, dng14@calstatela.edu † Department of Civil Engineering, California State University, Los Angeles Los Angeles, CA, USA jlucey2@calstatela.edu Abstract—Unmanned Aerial Vehicle (UAV) swarms increas- ingly support safety-critical applications that rely on distributed visual perception. Meeting the low-latency requirements of these applications can require perception models to execute within the swarm on inference-capable UAVs, creating a need for efficient UAV-to-UAV transport of high-bandwidth perception data. However, the Quality-of-Service (QoS) requirements of perception differ from conventional packet-level QoS; successful delivery of individual packets does not ensure that a complete, timely, and usable image is available for inference. We present a novel perception-aware communication middleware that treats complete perception-data samples as the communication objects for which QoS must be satisfied. The middleware extends a lightweight UDP broker-based publish-subscribe architecture with perception-specific services, including image fragmentation and reconstruction, concurrent packet transmission, priority- aware scheduling, and image quality assessment. The middleware is evaluated on a heterogeneous hardware testbed emulating a UAV swarm using YOLOv8n object detection. Experimental results demonstrate low end-to-end application latency, substan- tially higher throughput than a lightweight UDP broker, effective prioritization of perception traffic under increasing background load, and mitigation of object-detection degradation through middleware-level image quality assessment. This work provides an initial framework for integrating AI-specific data handling into communication middleware to support emerging distributed AI applications in multi-agent mobile cyber-physical systems. Index Terms—UAV swarms, perception-aware communication, communication middleware, edge AI, visual perception, QoS I. INTRODUCTION Unmanned Aerial Vehicle (UAV) swarms, defined as col- laborative and cooperative group of UAVs working to meet a common objective, can be utilized to support numerous safety-critical and time-sensitive applications in critical in- frastructure domains. Many of these applications, e.g. post- disaster search and rescue (SAR) and battlefield reconnais- sance, are driven by continuous video sensing to provide real- time situational awareness [1]. These applications increasingly rely on compute- and memory- intensive deep learning-based perception models, such as object detection, to achieve mission This work was supported by the Provost SEED Grant at California State University, Los Angeles. Manveen Kaur and Kevin Loi contributed equally to this work. Manveen Kaur is the corresponding author (email: mkaur39@calstatela.edu). objectives [2], [3]. Further, safety-critical applications must also operate under stringent Quality of Service (QoS) guaran- tees for latency and reliability, and in dynamic environments that frequently lack reliable communication infrastructure or low-latency access to remote computational resources [2]. Because UAV swarms are typically heterogeneous, with most UAVs constrained in compute, memory, and commu- nication resources [4], perception applications may require distributed operation where image and video sensing are de- coupled from perception-model execution. Furthermore, image and video sensing generate high-bandwidth perception data whose timely and reliable delivery directly influences the perception models’ inference accuracy and overall application performance. Therefore, UAV swarms must support the trans- port and processing of high-bandwidth perception data while being mindful of strict QoS requirements. Conventional ap- proaches for incorporating perception models in UAV swarms involve running model inference on remote cloud or edge computing infrastructure [4]. However, UAV swarms deployed in disaster response, environmental monitoring, or battlefield operations cannot always rely on persistent or timely connec- tivity to such infrastructure. Integrating Edge AI [5], i.e., running AI models on local devices, with Mobile Edge Computing, i.e., utilizing compu- tationally capable UAVs as mobile edge nodes, enables near- real-time onboard perception by executing perception models locally within the swarm. Here, UAVs can be designated as Edge AI UAVs or Sensing UAVs, with UAV-to-UAV communi- cation supporting perception-data exchange between the two. Three technological drivers underpin this vision: (1) advances in compact AI accelerator-equipped System-on-Chips (SoCs) that enable Edge AI [6], (2) the design of lightweight and dis- tributed AI/ML models for resource-constrained multi-agent systems [7], [8], and (3) the development of perception-aware communication middleware that extends QoS support beyond individual packets to complete data samples required for inference. Unlike conventional packet-centric communication, this middleware should identify, reconstruct, and prioritize complete perception inputs to meet application-level QoS requirements. This distinction is critical because successful delivery of individual packets does not guarantee usable arXiv:2609.24964v1 [cs.MA] 21 Sep 2026 perception data; fragments from multiple incomplete images may satisfy packet-level QoS requirements while providing no complete image suitable for inference. While recent research has demonstrated substantial progress toward the first two drivers through the emergence of numerous AI accelerator- equipped embedded platforms [9], [10] and model optimiza- tion techniques [7], distributed perception applications remain fundamentally dependent on the timely and reliable transport of high-bandwidth perception data among UAVs. Existing lightweight communication middleware, including MQTT, CoAP, and ZeroMQ, provide efficient message transport but remain unaware of data samples and the needs of perception models [11], [12]. In this work, we present a perception-aware communica- tion middleware for UAV swarms that treats complete im- age samples as the data units for which QoS requirements must be satisfied. The middleware extends a UDP broker- based publish-subscribe architecture with services tailored to distributed perception and transmission of complete images, including thread pooling for high-throughput transmission of image fragments, image-specific header fields to maintain semantic context over multiple image fragments, priority- aware scheduling, middleware-level image reconstruction, and image quality assessment to filter degraded inputs before infer- ence. The middleware is implemented in C as an application layer service, and evaluated on a heterogeneous multi-agent hardware testbed representative of resource-constrained UAV swarms. Experimental results demonstrate high-throughput perception-data transport with low end-to-end application la- tency, effective prioritization of QoS-sensitive image traffic under increasing network load, and filtering of degraded images to preserve downstream object-detection performance. The remainder of this paper is organized as follows. Sec- tion I presents the background and related work. Section I introduces the reference UAV swarm architecture and describes the salient features of the proposed communication middleware. Section IV details the experimental methodology and evaluation strategy. Section V presents and discusses the experimental results. Finally, Section VI concludes the paper. I. BACKGROUND AND RELATED WORKS Various technological advances are enabling distributed perception in resource-constrained UAV swarms. Advances in embedded edge AI hardware have substantially increased onboard computational capability, with modern System-on- Chips (SoCs) integrating dedicated AI accelerators, includ- ing Graphics Processing Units (GPUs), application-specific integrated circuits (ASICs), and field-programmable gate ar- rays (FPGAs), to accelerate deep learning inference. Exam- ples include the Qualcomm DragonWing [9] and NVIDIA Jetson families [10]. In parallel, model optimization tech- niques such as pruning, quantization, and knowledge distil- lation reduce model size, memory consumption, and inference latency [7], while inference frameworks such as NVIDIA TensorRT and TensorFlow Lite further optimize execution on embedded hardware. Distributed AI paradigms, including Federated Learning and Multi-Agent Reinforcement Learning, further enable AI workloads to be distributed across multiple agents [7], [8]. Deep learning-based visual perception has consequently be- come a fundamental capability for safety-critical autonomous UAV swarm applications. In SAR, representative approaches include using RGB cameras and a CNN-based object detector to identify survivors [3] and executing a pruned YOLOv5 model onboard each UAV for object detection [2]. However, many existing perception approaches assume either sufficient computational resources for local inference or reliable, low- latency, high-bandwidth connectivity to remote edge infras- tructure. While some works explicitly consider communication constraints in distributed UAV perception, primarily by re- ducing communication overhead through selective information exchange, optimizing communication resources, or improving network connectivity [13], [14], comparatively less attention has been given to communication middleware that manages perception data itself and its end-to-end transport between sensing and edge AI UAVs. Existing communication middleware, including MQTT, CoAP, and ZeroMQ, provides efficient and lightweight mes- sage transport for distributed systems but is largely ap- plication agnostic. MQTT provides broker-based publish- subscribe messaging with multiple QoS levels; CoAP pro- vides lightweight RESTful communication, typically over UDP, for constrained devices; and ZeroMQ provides high- performance asynchronous messaging patterns for distributed applications [11], [12]. While suitable for general-purpose communication, these technologies do not natively manage perception data as complete communication units, includ- ing their fragmentation and reconstruction or preservation of application-level priority across image fragments. Emerging technologies such as Multi-access Edge Computing (MEC) and 5G NR Sidelink provide complementary capabilities for low-latency edge computing and direct device-to-device com- munication, but depend on compatible network infrastructure and hardware [15]. Thus, a gap remains for perception-aware communication middleware that bridges communication and perception by managing complete perception-data units while satisfying the latency, reliability, and bandwidth requirements of distributed edge inference. I. SOLUTION DESIGN Swarm Configuration: The proposed communication mid- dleware assumes a distributed, decentralized, hierarchical UAV swarm architecture that supports scalable coordination among heterogeneous UAVs operating in communication-constrained environments [16]. As shown in Fig. 1, the swarm is parti- tioned into multiple heterogeneous clusters, each managed by a cluster manager responsible for coordinating local cluster operations. A swarm leader oversees inter-cluster coordination by forming a top-level cluster with the cluster managers, facilitating information exchange among clusters. Tasks are assigned according to UAV capabilities. Resource-constrained follower UAVs primarily perform sensing and data collec- tion, while computationally capable cluster managers and the swarm leader coordinate task execution. Depending on mission requirements, one or more computationally capable UAVs within the swarm may serve as mobile edge AI nodes, executing perception models on behalf of resource-constrained sensing UAVs. To improve robustness and eliminate a single point of failure, the roles of swarm coordination and edge AI execution are decoupled; thus, swarm leaders may not function as mobile edge AI nodes. It is not necessary for each cluster to have its own edge AI node(s). While the proposed middleware can support all inter-UAV communication, including control, coordination, and application data exchange, the remainder of this paper focuses exclusively on the communication support- ing distributed visual perception. Swarm Leader Cluster Manager Cluster Followers Camera Sensor equipped UAV Camera Sensor equipped UAV Camera Sensor equipped UAV Mobile Edge AI UAV CLUSTER 1 CLUSTER 2 CLUSTER 3 CLUSTER 0 Lateral View of UAV Swarm Top Layer (Swarm Leader) Middle Layer (Cluster Managers) Bottom Layer (Cluster Followers) Leader - Cluster manager Communication Intra-cluster Communication Legend Top-View of UAV Swarm Experimental Design Fig. 1: Proposed UAV Swarm Configuration and Experimental Testbed Configuration (bottom right) Proposed Communication Middleware: The proposed middleware extends a lightweight UDP broker-based publish- subscribe architecture and comprises two software modules: a client module, which provides publisher and subscriber functions on follower and edge AI UAVs, and a broker module, which executes on cluster managers and swarm leaders to facilitate communication among clients. The swarm hierarchy determines the broker servicing each client, while inter-broker communication is supported through broker-to-broker sub- scriptions. As the underlying publish-subscribe mechanisms, including topic-based publication, subscription, and broker operation, are well established, we refer readers to prior literature for further details. The middleware supports two communication modes: best-effort, in which messages are transmitted without acknowledgment, and reliable, in which published messages require acknowledgment, analogous to MQTT QoS 0 and QoS 1, respectively. This work evaluates the best-effort mode. Unlike conventional packet-centric communication, the middleware treats perception data as the fundamental com- munication object and provides the following core services: • Image Fragmentation and Reconstruction: Standard- resolution images exceed the maximum payload supported by a network packet, requiring fragmentation before transmission. The client code fragments each sensing image, and adds semantic context through header fields, such as image id, fragmentid, and totalfragments. The client code com- putes the total fragments, F, in an image as S/P, where S is the file size and P is the maximum payload size for each packet. Each fragment is assigned a msg id, imageid, and fragment id prior to transmission. Zero-padding is used if a fragment size is ≤ P. Upon receiving image fragments, the image reconstruction involves writing each fragment into a new file in sequence (using fragment id). Image fragmen- tation and reconstruction preserve the image file format. At the receiver, the client module maintains an image log that tracks the reconstruction status of each partially received image. Once all fragments of an image are received and reconstruction is complete, its entry is removed from the log. The client waits for a predefined timeout period for all fragments to arrive; if the timeout expires, the image is reconstructed using the available fragments. The timeout is dynamically determined using an Exponentially Weighted Moving Average (EWMA) of the measured round-trip time (RTT). • Concurrent Packet Transmission: Conventional sequential processing and transmission of image fragments can limit achievable throughput and increase the time required to de- liver a complete image to the perception model. To improve throughput and reduce image delivery latency, the proposed middleware employs concurrent packet transmission using a fixed-size thread pool in both client and broker modules. The thread pool supports two task types, image fragmen- tation and packet transmission, submitted to a shared job queue. Fragmentation tasks divide an image concurrently by assigning distinct image regions to available worker threads; once fragmentation is complete, transmission tasks publish the resulting fragments through UDP sockets to the broker or subscriber. Mutex-protected critical sections ensure thread- safe access to shared buffers. The number of worker threads is configurable, balancing communication performance against processor utilization; larger pools increase concurrency and can reduce image delivery latency at the cost of greater CPU utilization, while smaller pools reduce computational overhead but may increase delivery time. • Priority-Aware Packet Scheduling: A client assigns a priority level to every transmitted packet, and the broker schedules outgoing messages using a multi-level circular ring-buffer priority queue. The scheduler maintains multiple independent circular buffers, one for each priority level. Incoming pack- ets are inserted into the tail of the buffer corresponding to their assigned priority, while transmission always services the highest non-empty priority queue first. A limitation of this priority scheduling is the possibility of starvation for lower- priority traffic during sustained bursts of high-priority image transmission. To mitigate this effect, the middleware employs priority aging. Each packet maintains a starvation counter that increases while waiting in the queue. When the counter ex- ceeds a predefined threshold determined by application needs, the packet’s priority is increased, ensuring that lower-priority packets are also serviced without significantly affecting latency for high-priority perception traffic. • Image Quality Assessment (IQA): Various factors can de- grade the quality of received perception data and reduce the performance of the perception model. Our current IQA imple- mentation focuses on image degradation due to noisy sensing data. To prevent degraded images from being forwarded to the perception model, the proposed middleware performs IQA fol- lowing image reconstruction using a computationally efficient sharpness metric called the Variance of the Laplacian which we implement natively in the middleware [17], [18]. The reconstructed image is converted to grayscale and processed using a standard four-neighbor discrete Laplacian operator to identify edges. The variance of the Laplacian response provides a sharpness score: high variance indicates strong edge information and a sharper image, while low variance indicates blur or degradation that may reduce inference reliability. The computed variance is compared against a configurable threshold selected through empirical calibration. Images whose sharpness score falls below the threshold are classified as low quality and are flagged. The current version of the middleware discards these images; however, it is possible to flag them for review by the perception models. IV. EXPERIMENT STRATEGY We consider an exemplar application of a heterogeneous UAV swarm supporting a latency-sensitive reconnaissance mission on the battlefield. The swarm is tasked with sur- veying an area of interest, and uses video sensing and a YOLOv8n perception model to identify objects of interest. Simplifying for validation, we use the swarm configuration shown in the bottom-right insert in Fig. 1. Here, one camera- sensor equipped follower UAV (U S ) conducts video-sensing, and a pre-trained YOLOv8n model runs on the Mobile Edge AI UAV (U Y v8 ). The client software runs on U S and U Y v8 , while the broker software runs on the swarm leader (U L ). Communication between U S and U Y v8 follows the path U S ↔ U L ↔ U Y v8 . We assume that the outcome of the perception model is utilized as an input for future path planning of the swarm. Although more recent YOLO variants offer improved computational efficiency, YOLOv8n was selected as the mature and stable release available during middleware development. Since the proposed middleware is independent of the underlying perception model, the choice of detector does not affect its communication mechanisms or the conclusions of this work. We train YOLOv8n using the AU-AIR dataset [19], which contains over 32,000 annotated RGB image frames extracted from low-altitude UAV videos across eight traffic-related object classes. The experimental perception task focuses on car detection. Model training was performed on a computationally-capable machine, and the pre- trained model was deployed on U Y v8 . Hardware Testbed: The proposed communication middle- ware was evaluated using a heterogeneous hardware testbed emulating the experimental configuration in Fig. 1. A Rasp- berry Pi 4 (RPi4), Intel Next Unit of Computing (NUC), and NVIDIA Jetson AGX Orin (Orin) serve as U S , U L , and U Y v8 , respectively. The RPi4 represents a resource-constrained UAV, while the NUC represents a comparatively more com- putationally capable UAV. The Orin serves as the edge AI node, providing dedicated hardware for neural-network in- ference not available on the RPi4 or NUC. All nodes are interconnected via wired Ethernet operating at 100 Mbps to isolate middleware performance from wireless effects such as channel contention, scheduling, and packet collisions, as well as bandwidth bottlenecks. Similarly, node mobility is not considered in this setup to avoid complications arising from network disruptions. System clocks are synchronized using the Chrony implementation of the Network Time Protocol (NTP). Experimental Design: We conducted three experiments. Experiment 1 transmits a set of 182 images from the AU- AIR dataset (unseen during YOLOv8n training) from U S to U Y v8 to evaluate network-level communication, middle- ware, and application performance. We measure end-to-end communication latency as the time between transmission and reception of each image fragment; middleware latency as the time between each fragment’s transmission until completion of image reconstruction and IQA; and application latency as the time between transmission of the first image fragment and completion of YOLOv8n inference. Corresponding packet-, middleware-, and application-level throughput are also mea- sured. Experiment 2 evaluates the middleware’s priority-aware packet scheduling under increasing network load. Alongside transmitting the 182-image dataset from U S to U Y v8 , a second sender (U ′ S ) simultaneously transmits equivalent volume of data to a second receiver (U ′ R ) through the same broker. Image fragments originating from U S are assigned a higher priority level (10) than background traffic (priority level 1). Four traffic profiles are evaluated by increasing the transmission rate of U ′ S from 1× to 4× that of U S , and the resulting end-to-end communication latency is measured. Experiment 3 evaluates the middleware’s IQA feature. To emulate degraded sensing conditions, 20, 40, and 60 of the 182 transmitted images are artificially degraded using Gaussian blur prior to transmission. Gaussian blur reduces image sharpness by averaging neighboring pixel values, with the blur radius controlling the degree of degradation without altering image resolution. Fig. 2 provides an illustration of image degradation used. The degraded images are interspersed with unmodified images and transmitted from U S to U Y v8 . The feature is evaluated on the middleware’s ability to filter degraded images and mitigate their impact on YOLOv8n object detection using the standard metrics of precision, recall, mAP@50, and mAP@50–95, whose definitions are omitted for brevity. V. RESULTS AND DISCUSSION Fig. 3a compares communication latency and throughput of the proposed middleware against a lightweight UDP broker for transmission of 10,000 packets between a publisher and sub- scriber in the hardware testbed. The UDP broker implements only basic list-based topic matching, without QoS mechanisms Fig. 2: Illustration of a Standard Image and its Degraded Version Created using Gaussian Blur. or advanced topic lookup. While the additional functionality and processing introduced by the proposed middleware results in approximately 40% higher per-packet latency than the UDP broker, threadpooling in the proposed middleware allows mul- tiple packet transmission operations to progress concurrently, increasing the observed throughput by more than 200%. The proposed middleware trades a smaller increase in individual packet latency for substantially higher aggregate data-transfer throughput, enabling more effective utilization of available communication capacity for high-bandwidth perception data. Fig. 3b shows communication and middleware latency ob- served in Experiment 1. The average per-packet communi- cation latency is 3.175 ms, with 90th- and 95th-percentile latencies of 3.0 ms and 20.49 ms, respectively. Thus, 90% of the 10,000 transmitted image fragments were delivered within 3.0 ms, while the substantially higher 95th-percentile value indicates a small number of significant tail-latency events. The higher tail-end latency values are attributed to the fact that, in the current implementation, multiple worker threads process image fragments and issue transmissions within a short interval, producing synchronized packet bursts that can increase contention at socket and kernel buffers. Thread man- agement will be further optimized in the next iteration of the middleware. The average per-packet middleware latency was 6.701 ms, approximately 2.1× the average communication latency, with 90th- and 95th-percentile values of 10.42 ms and 20.54 ms, respectively. The increase results from image reconstruction and IQA steps. Ablation experiments with IQA disabled showed computationally-complex Variance-of- Laplacian quality assessment accounts for ≈66% of the addi- tional processing overhead. Fig. 3c shows the per-image application latency observed in Experiment 1. Each experimental image comprised approx- imately 55 fragments on average. The resulting average appli- cation latency was 21.23 ms, with 90th- and 95th-percentile values of 32.76 ms and 40.38 ms, respectively. Concurrent fragment processing and transmission allow the latency of these fragments to overlap rather than accumulate sequentially. Thus, the proposed middleware enables an image to progress from transmission at U S , through reconstruction and IQA, to completion of object detection at U Y v8 in 21.23 ms on aver- age, with 95% of images completing the end-to-end perception pipeline within 40.38 ms. Communication and middleware throughput observed in Experiment 1 averaged 4.24 Mbps and 2.22 Mbps, respectively, over experiment duration. The reduction in observed middleware throughput is consistent TABLE I: Impact of IQA on Object Detection Performance. Images BlurredPrecisionRecallmAP@50mAP@50–95 Image Quality Assessment: Off 00.7120.8010.6450.394 200.6250.6880.5280.313 400.6280.6240.5170.308 600.6320.5480.4790.281 Image Quality Assessment: On 00.7120.8010.6450.394 200.6880.7880.6230.368 400.7150.8380.7190.428 600.7390.8870.8070.472 with increase in latency from the additional reconstruction and IQA processing time. Fig. 4 shows the communication latency of high-priority image packets under increasing lower-priority background traffic. Traffic Profiles 0–3 correspond to U ′ S transmitting similarly sized lower-priority packets at 1×, 2×, 3×, and 4× the transmission rate of U S , respectively, while U S transmitted with a higher-priority. Despite increasing background traffic from 1× to 4×, the tail latency of high-priority image packets increased only marginally, by approximately 1 ms across the evaluated traffic profiles. These results provide an initial val- idation that the broker’s priority-aware scheduler can identify and preferentially service high-priority perception traffic under increasing load. A more comprehensive evaluation, including the resulting latency and potential starvation of competing lower-priority traffic, is left for future work. Table I illustrates the middleware’s IQA capability as demonstrated through impact on YOLOv8n’s object detection performance. Results are shown for the car class, which is sufficient to evaluate the impact of degraded perception data on the target application. Overall detection performance is relatively modest. This is consistent with the class im- balance and baseline performance reported for the AU-AIR dataset [19]. With IQA disabled, introducing 20 blurred images reduces precision, recall, mAP@50, and mAP@50–95 by 12.22%, 14.11%, 18.14%, and 20.56%, respectively. With increasing degradation, recall decreases from 0.801 to 0.548 and mAP@50 from 0.645 to 0.479, while precision stabilizes near 0.63 after its initial decline. This suggests that further degradation primarily increases missed detections rather than false positives. With IQA enabled, the corresponding reduc- tions at 20 blurred images are limited to 3.37%, 1.62%, 3.41%, and 6.60%. Metrics subsequently increase at 40 and 60 blurred images because filtering changes the composition of images reaching YOLOv8n, creating a selection effect rather than improving the detector itself. Thus, IQA prevents unsuitable perception data from reaching YOLOv8n and substantially mitigates degradation in object detection performance. VI. CONCLUSION AND FUTURE WORK In this work, we implemented and validated an initial version of a perception-aware communication middleware for distributed perception in heterogeneous UAV swarms. The middleware achieves low end-to-end communication and application latency while substantially improving through- Comparison of Latency and Throughput Observed using the Proposed Middleware vs. a Simple UDP Broker-based Publish/Subscribe Implementation (a) Proposed Middleware vs. a Simple UDP Broker-based Implementation. Probability Density Latency (ms) Communication Latency Middleware Latency Probability Distribution of Communication and Middleware Latency over Experiment Duration 90th Percentile Value for Communication Latency = 3.0 ms 90th Percentile Value for Middleware Latency = 10.42 ms 95th Percentile Value for Communication Latency = 20.49 ms 95th Percentile Value for Middleware Latency = 20.54 ms (b) End-to-End Communication and Middle- ware Latency. Probability Distribution of Application Latency over Experiment Duration Probability Density Latency (ms) Application Latency 90th Percentile Value for Application Latency = 32.76 ms 95th Percentile Value for Application Latency = 40.38 ms (c) End-to-End Application Latency. Fig. 3: Comparison of the Middleware with a Simple UDP broker and Results from Experiment 1. Fig. 4: Latency Observed for High-Priority Image Packets under Varying Background Traffic Conditions. put over a lightweight UDP broker baseline. Priority-aware scheduling maintains the latency of perception traffic under increasing background load, while middleware-level image quality assessment mitigates degradation in object detection performance caused by low-quality sensing data. Collectively, these results demonstrate the value of integrating perception awareness directly into the communication middleware. Future work will focus on iterative improvements and broader valida- tion of the middleware. Thread management will be optimized to reduce tail latency, and priority-aware scheduling will be evaluated more extensively, particularly its impact on lower- priority traffic. The overarching goal is to package and release the middleware as an open-source platform for implementing and evaluating distributed perception applications in heteroge- neous UAV swarms. REFERENCES [1] J. Zhao, Q. Li, P. Chi, and Y. Wang, “Active vision-based uav swarm with limited fov flocking in communication-denied scenarios,” IEEE Transactions on Instrumentation and Measurement, 2025. [2] S. Moon, J. Jeon, D. Kim, and Y. Kim, “Swarm reconnaissance drone system for real-time object detection over a large area,” IEEE Access, vol. 11, p. 23 505–23 516, 2023. [3] J. Horyna, T. Baca, V. Walter, D. Albani, D. Hert, E. Ferrante, and M. Saska, “Decentralized swarms of unmanned aerial vehicles for search and rescue operations without explicit communication,” Autonomous Robots, vol. 47, no. 1, p. 77–93, 2023. [4] Y. Alqudsi and M. Makaraci, “Uav swarms: research, challenges, and future directions,” Journal of Engineering and Applied Science, vol. 72, no. 1, p. 12, 2025. [5] R. Singh and S. S. Gill, “Edge ai: a survey,” Internet of Things and Cyber-Physical Systems, vol. 3, p. 71–92, 2023. [6] A. Reuther, P. Michaleas, M. Jones, V. Gadepally, S. Samsi, and J. Kepner, “Ai and ml accelerator survey and trends,” in 2022 IEEE High Performance Extreme Computing Conference (HPEC).IEEE, 2022, p. 1–10. [7] W. Su, L. Li, F. Liu, M. He, and X. Liang, “Ai on the edge: a comprehensive review,” Artificial Intelligence Review, vol. 55, no. 8, p. 6125–6183, 2022. [8] Y. Ding, Z. Yang, Q.-V. Pham, Y. Hu, Z. Zhang, and M. Shikh-Bahaei, “Distributed machine learning for uav swarms: Computing, sensing, and semantics,” IEEE Internet of Things Journal, vol. 11, no. 5, p. 7447– 7473, 2023. [9] QualcommTechnologies,Inc.,“QualcommDragonwingIQ8 SeriesPlatformProductBrief,”QualcommTechnologies, Inc.,Tech.Rep.,2026,accessed:2026-07-27.[Online]. Available: https://docs.qualcomm.com/bundle/publicresource/87-83839- 1 REVBQualcommDragonwingIQ8SeriesPlatformProductBrief.pdf [10] NVIDIACorporation,“NVIDIAJetsonAGXOrinSeries TechnicalBrief,”NVIDIACorporation,Tech.Rep.,2025, version1.2.Accessed:2026-07-27.[Online].Available: https://w.nvidia.com/content/dam/en-z/Solutions/gtcf21/jetson- orin/nvidia-jetson-agx-orin-technical-brief.pdf [11] N. Naik, “Choice of effective messaging protocols for iot systems: Mqtt, coap, amqp and http,” in 2017 IEEE international systems engineering symposium (ISSE). IEEE, 2017, p. 1–7. [12] Y. Liu and N. Ahmed, “Performance evaluation of mqtt and zeromq messaging protocols for uav applications,” in 2025 12th International Conference on Wireless Networks and Mobile Communications (WIN- COM). IEEE, 2025, p. 1–6. [13] Y. Hu, S. Fang, Z. Lei, Y. Zhong, and S. Chen, “Where2comm: Communication-efficient collaborative perception via spatial confidence maps,” Advances in neural information processing systems, vol. 35, p. 4874–4886, 2022. [14] N. Mehrabi, S. P. H. Boroujeni, J. Hofseth, A. Razi, L. Cheng, M. Kaur, J. Martin, and R. Amin, “Adaptive data transport mechanism for uav surveillance missions in lossy environments,” in 2025 IEEE 22nd Consumer Communications & Networking Conference (CCNC). IEEE, 2025, p. 1–4. [15] H. Bagheri, M. Noor-A-Rahim, Z. Liu, H. Lee, D. Pesch, K. Moessner, and P. Xiao, “5g nr-v2x: Toward connected and cooperative autonomous driving,” IEEE Communications Standards Magazine, vol. 5, no. 1, p. 48–54, 2021. [16] M. Kaur, A. Razi, L. Cheng, R. Amin, and J. Martin, “Design and evalu- ation of an application-oriented data-centric communication framework for emerging cyber-physical systems,” in 2023 IEEE 20th Consumer Communications & Networking Conference (CCNC). IEEE, 2023, p. 875–878. [17] S. Pertuz, D. Puig, and M. A. Garcia, “Analysis of focus measure operators for shape-from-focus,” Pattern Recognition, vol. 46, no. 5, p. 1415–1432, 2013. [18] B. S ̧ ims ̧ek and H. S ̧ . Bilge, “A novel motion blur resistant vslam framework for micro/nano-uavs,” Drones, vol. 5, no. 4, p. 121, 2021. [19] I. Bozcan and E. Kayacan, “Au-air: A multi-modal unmanned aerial vehicle dataset for low altitude traffic surveillance,” in 2020 IEEE international conference on robotics and automation (ICRA).IEEE, 2020, p. 8504–8510.