Performance evaluation of recent object detection models for traffic safety applications on edge

Abstract

Real-time objection detection is becoming more important and critical in all application areas, including Smart Transport and Smart City. From safety/security to resource efficiency, real-time image processing approaches are used more than ever. On the other hand, low-latency requirements and available resources present challenges. Edge computing integrated with cloud computing minimizes communication delays but requires efficient use of resources due to its limited resources. For example, although deep learning-based object detection methods give very accurate and reliable results, they require high computational power. This overhead reveals a need to implement deep learning models with less complex architectures for edge deployment. In this paper, the performance of evolving deep learning models with their lightweight versions such as YOLOv5-Nano, YOLOX-Nano, YOLOX-Tiny, YOLOv6-Nano, YOLOv6-Tiny, and YOLOv7-Tiny are evaluated on a commercially available edge device. The results show that YOLOv5-Nano and YOLOv6-Nano with their TensorRT versions can provide real-time applicability in approximately 35 milliseconds of inference time. It is also observed that YOLOv6-Tiny gives the highest average precision while YOLOv5-Nano gives the lowest energy consumption when compared to other models.