Source-linked AI summary
A Performance Analysis of You Only Look Once Models for Deployment on Constrained Computational Edge Devices in Drone Applications
Lucas Rey, Ana M. Bernardos, Andrzej D. Dobrzycki, David Carramiñana, Luca Bergesio, Juan A. Besada, José Ramón Casar
TL;DR
Onboard drone AI remains constrained by limited computation, while evidence on integrated edge–cloud latency and current YOLO deployments remains limited. This study evaluates YOLOv8 models, quantization, and edge–cloud processing, finding trade-offs among speed, accuracy, stability, energy efficiency, and communication latency.
Problem
Onboard drone AI requires autonomous, energy-efficient object detection, but studies provide limited evidence on integrated edge–cloud latency and hardware–software interactions.
Method
The study combines isolated device experiments with a representative drone image-processing pipeline to evaluate inference speed, power consumption, quantization, and edge versus cloud deployment.
Results
YOLOv8 quantization and deployment produced device-dependent trade-offs: INT8 was fastest but reduced accuracy, Orin devices were more stable than Raspberry Pi 5, and cloud communication latency limited real-time viability.
Takeaways & Limitations
Real-time UAV deployment requires balancing detection precision, inference speed, stability, energy efficiency, and communication latency, with hybrid architectures offering a supported optimization strategy.
Takeaways & Limitations
Raspberry Pi 5 testing was incomplete because computational limitations caused prediction latency, instability, and unreliable operation in the integrated architecture.
Abstract
from arXiv · showhide
Advancements in embedded systems and Artificial Intelligence (AI) have enhanced the capabilities of Unmanned Aircraft Vehicles (UAVs) in computer vision. However, the integration of AI techniques o-nboard drones is constrained by their processing capabilities. In this sense, this study evaluates the deployment of object detection models (YOLOv8n and YOLOv8s) on both resource-constrained edge devices and cloud environments. The objective is to carry out a comparative performance analysis using a representative real-time UAV image processing pipeline. Specifically, the NVIDIA Jetson Orin Nano, Orin NX, and Raspberry Pi 5 (RPI5) devices have been tested to measure their detection accuracy, inference speed, and energy consumption, and the effects of post-training quantization (PTQ). The results show that YOLOv8n surpasses YOLOv8s in its inference speed, achieving 52 FPS on the Jetson Orin NX and 65 fps with INT8 quantization. Conversely, the RPI5 failed to satisfy the real-time processing needs in spite of its suitability for low-energy consumption applications. An analysis of both the cloud-based and edge-based end-to-end processing times showed that increased communication latencies hindered real-time applications, revealing trade-offs between edge (low latency) and cloud processing (quick processing). Overall, these findings contribute to providing recommendations and optimization strategies for the deployment of AI models on UAVs.
1. Introduction
UAV computer vision can improve autonomous detection and tracking, but onboard deployment is constrained by processing and energy limits. The study evaluates YOLOv8 variants, quantization, and embedded platforms for real-time drone applications.
- Onboard AI enables drones to operate autonomously in low-connectivity settings such as remote maintenance, defense, and emergency response.
- YOLOv8n and YOLOv8s use fewer parameters than larger variants, trading some accuracy for computational efficiency on constrained edge devices.
- Quantization reduces bit width, memory footprint, and computational load, supporting more efficient execution on limited hardware.
- The study evaluates YOLOv8 across Jetson Orin Nano, Jetson Orin NX, and Raspberry Pi 5 for real-time detection suitability.
- The evaluation examines accuracy, processing speed, power consumption, quantization, and representative drone image-processing pipelines.
2. The State of the Art
Prior work studies edge-device optimization, quantization, and distributed edge–cloud processing, but evidence remains limited for integrated real-time YOLOv8 drone deployments. This study addresses gaps in cross-platform accuracy, latency, energy, and hardware–software interaction.
- 2.1. Deploying Deep Learning on Energy-Efficient Devices: Research benchmarks edge hardware and deep-learning configurations to balance inference speed, energy consumption, and real-time processing.
- 2.1. Deploying Deep Learning on Energy-Efficient Devices: Optimization frameworks combine pruning, quantization, and runtime techniques to improve inference efficiency across high-end and low-end edge devices.
- 2.1. Deploying Deep Learning on Energy-Efficient Devices: Distributed approaches such as DistrEdge and Edge YOLO distribute inference across devices or edge–cloud systems to improve load balancing and latency.
- 2.1. Deploying Deep Learning on Energy-Efficient Devices: Hardware-specific frameworks matter: TF-TRT improves Jetson inference, whereas TF-Lite does not use Jetson GPU resources effectively.
- 2.2. Quantization Techniques in Object Detection Models: PTQ, QAT, and mixed-precision quantization reduce computational and memory demands, with QAT sometimes achieving higher accuracy than PTQ.
- 2.2. Quantization Techniques in Object Detection Models: Evidence remains limited for quantization effects on recent YOLO models in real-time drone applications and for latency between hardware and software processes.
- 2.2. Quantization Techniques in Object Detection Models: This study focuses on YOLOv8 accuracy, inference speed, and energy consumption to derive configuration recommendations for constrained Jetson and Raspberry Pi platforms.
3. Methodology
The methodology evaluates YOLOv8 models across heterogeneous edge devices using annotated aerial imagery, deployment-specific optimization, and isolated and pipeline-level experiments. It compares model efficiency, accuracy, inference speed, and energy-related trade-offs under different precision formats.
- Hardware Platforms: The study evaluates YOLOv8 models and quantized versions on Raspberry Pi 5, Jetson Orin Nano, and Jetson Orin NX platforms with different hardware capabilities.The devices are tested under uniform experimental conditions to assess object detection under varied processing constraints.
- Evaluation Strategy: The evaluation combines isolated inference tests with deployment in a representative drone image-processing pipeline.Experiments measure inference speed and power consumption in isolation, then compare edge and cloud deployment in a realistic pipeline.
- Dataset Creation: The dataset contains 6000 manually annotated 640 × 640 aerial images captured across varied drone positions, trajectories, altitudes, and viewing angles.DJI Tello drones and TurtleBot ground robots were used in an indoor environment with varied furniture and OptiTrack positioning.
- Model Selection: YOLOv8n and YOLOv8s were selected because their reduced depth and parameter counts target computationally constrained edge devices.YOLOv8n prioritizes speed, whereas YOLOv8s increases computational demand to improve detection accuracy.
- Edge Optimization: The models were optimized using FP16 and INT8 quantization, with TensorRT on Jetson devices and NCNN on Raspberry Pi 5.Jetson platforms were evaluated at FP32, FP16, and INT8, while Raspberry Pi 5 used FP32 through NCNN.
4. Performance Evaluation of Quantized YOLO over Resource-Constrained Devices
The evaluation compares quantized YOLOv8s and YOLOv8n across Jetson Orin and Raspberry Pi devices, showing that Orin platforms provide substantially faster inference while energy-performance trade-offs vary by configuration.
- Inference performance: 15.16 ms was the shortest mean iteration time, achieved by YOLOv8n_INT8 on the Jetson Orin NX.On the Orin Nano, YOLOv8n_INT8 achieved 23.16 ms, while Raspberry Pi inference was substantially slower.
- Quantization constraints: FP16 and INT8 were not feasible on Raspberry Pi 5 because NCNN precision support was limited.Only the stable FP32 configuration was included in the comparative analysis; attempted reduced-precision configurations showed no performance improvement and a notable accuracy loss.
- Energy consumption: 10–14 W was the Orin NX power-consumption range, compared with 7.4–8.7 W for the Orin Nano.The Orin NX required more power but delivered notable inference-speed improvements, particularly with INT8 quantization.
- Accuracy and speed: 37.90 FPS and mAP50-95 of 0.8622 were achieved by YOLOv8s with FP16 quantization on the Orin Nano.This slightly exceeded YOLOv8s with FP32 on the Orin NX, which achieved 35.65 FPS and mAP50-95 of 0.8610.
- Accuracy and speed: 7.32 FPS was achieved by YOLOv8s_FP32 on the Raspberry Pi 5, with mAP50-95 of 0.8620.The Raspberry Pi used FP32 through NCNN and was characterized as better suited to less demanding applications where real-time processing is not critical.
- Energy consumption: 1.498 J/inference was measured for Raspberry Pi YOLOv8s FP32, despite its lowest absolute power consumption.Its low prediction rate reduced energy efficiency relative to the Jetson devices.
5. Performance Evaluation Using a Realistic Testbed
The realistic UAV testbed routes video from a ground control station to either edge or cloud inference, enabling comparisons of device latency and deployment trade-offs. Edge processing reduced communication delay, while cloud processing provided faster model computation but much higher end-to-end latency.
- Deployment architecture: Video frames were forwarded from the ground control station to either a TensorRT-optimized edge device or a cloud instance for object detection.The architecture used WebSocket communication and measured video transmission, detection, and prediction-distribution phases.
- Edge deployment: 29.95 ms was the minimum Orin NX RTT, achieved with YOLOv8n INT8.The Orin Nano minimum RTT was 44.27 ms with YOLOv8s INT8, while the Raspberry Pi test was incomplete because of computational limitations and prediction latency.
- Edge deployment: 14.29 ms was the minimum processing time on the Orin NX, achieved with YOLOv8n INT8.The Orin Nano minimum was 20.54 ms with the same model and quantization.
- Edge–cloud comparison: 35.09 ms was the mean edge RTT versus 348.21 ms for cloud deployment using YOLOv8s FP16.Mean communication latency was 2.50 ms at the edge and 341.41 ms in the cloud.
- Edge–cloud comparison: 6.82 ms was the cloud processing latency versus 32.59 ms at the edge.The cloud’s computational advantage was counteracted by its much higher communication latency, resulting in performance constraints.
6. Discussion
The evaluation identifies device, quantization, and processing-location trade-offs for real-time drone vision. Orin platforms provide stronger and more stable edge performance, while cloud communication latency and Raspberry Pi limitations constrain real-time use.
- Quantization trade-offs: 41.20 FPS with YOLOv8s INT8 on Orin Nano exceeded 27.00 FPS with FP32, while mAP50-95 fell from 0.8608 to 0.7968.FP16 provided 37.90 FPS and 0.8622 mAP50-95, offering a balance between speed and accuracy.
- Quantization trade-offs: 48.24 FPS and 0.8620 mAP50-95 with YOLOv8s_FP16 on Orin NX exceeded YOLOv8n_FP32 at 35.65 FPS and 0.8420 mAP50-95.
- Device performance: Orin models outperformed the Raspberry Pi 5 in real-world-like architectures, although Orin Nano processing still incurred frame-capture and internal-operation overhead.Improved data capture and data-flow management could reduce this overhead.
- Device performance: The Orin NX showed low response-time variability and strong stability, whereas Raspberry Pi 5 interruptions and control issues limited complex object-detection architectures.The Raspberry Pi 5 is therefore more appropriate for basic processing or simpler algorithms.
- System trade-offs: Jetson devices combined stronger computational performance with higher energy use, while Raspberry Pi 5 offered energy efficiency but struggled with real-time detection.The authors propose assigning lightweight tasks to Raspberry Pi and critical detection tasks to Jetson devices.
- Edge–cloud processing: Cloud processing had low model latency but high communication latency, whereas GPU-equipped edge devices enabled rapid local processing for real-time drone applications.A closer AWS region may improve cloud results, but communication bottlenecks remain.
- Operational recommendations: Table 7 aligns Orin NX INT8, Orin Nano FP16, and Orin Nano INT8 configurations with extended missions, high-quality detection, and low-demand tasks, respectively.The guide selects hardware and quantization settings according to operational requirements such as coverage, tracking, and precision.
- Future scope: Outdoor variability, dynamic conditions, and emerging low-power hardware remain important areas for extending these controlled indoor findings.The paper specifically calls for outdoor validation and investigation of energy-efficient inference and FPGA or low-power-chip co-design.
7. Conclusions
The study provides practical guidance for deploying YOLOv8 models on constrained drone-computing devices by evaluating performance and configuration trade-offs. Its conclusions also identify indoor testing limits and future directions involving newer YOLO versions, hybrid architectures, and advanced networks.
- Conclusions: YOLOv8n and YOLOv8s can operate efficiently on Raspberry Pi 5, Orin NX, and Jetson Orin Nano, but INT8 speed gains may reduce detection accuracy.
- Conclusions: Recommendations link device configurations and quantization methods to operational demands, helping practitioners select settings for diverse edge-computing scenarios.
- Limitations: Indoor testing was controlled and repeatable but could not fully reproduce outdoor lighting variation, environmental interference, or multi-agent dynamics.More complex mock-ups are needed to assess reliability under conditions closer to real-world operations.
- Future work: Future work should compare YOLOv10 and YOLOv11 with YOLOv8 for accuracy, efficiency, and real-time performance in drone applications.
- Future work: Hybrid Jetson, Raspberry Pi, and cloud architectures warrant real-world testing for feasibility and effectiveness in energy-sensitive missions.
- Future work: Integrating edge architectures with 5G and tactical cloud technologies could support decentralized swarm processing and reduce data-transmission latency.Such systems would also require larger drones to accommodate added weight and power requirements.