Comparative Analysis of Lightweight Vision Models on Embedded Accelerator Devices

The rapid growth of edge devices demands efficient neural network accelerators for low-latency, energy-efficient AI tasks like object detection and pose estimation. Despite benefits such as improved privacy and reduced cloud dependence, edge AI faces challenges including hardware heterogeneity, resource constraints, and the balance between real-time performance and accuracy. This paper benchmarks STMicroelectronics STM32N6, Luxonis OAK-FCC 4P, and SONY IMX500 on object detection and human pose estimation tasks, evaluating power consumption, inference time, accuracy, and memory usage. Using YOLOv5 and YOLOv8 models in Small and Nano configurations on COCO and COCO-Pose datasets, the study highlights the trade-offs between efficiency and accuracy. Results show IMX500 leads in inference time and power efficiency but is limited in memory compared to N6 and OAK-FCC 4P. This study provides a comprehensive evaluation of the suitability of different hardware architectures, guiding the optimization of lightweight neural network deployment on resource-constrained platforms.