Performance Evaluation of CNN-based Object Detectors on Embedded Devices

Computer-vision algorithms have been used to enhance hands-free user interactions with smart systems. Traditional approaches for object and gesture detection and recognition are through the processing of video frames by convolutional neural networks (CNN) on cloud or edge servers. However, this approach limits the frame processing rate due to the latency incurred in the video frame offloading process and communication to distant cloud and congested edge servers. Thus, a promising approach is to perform part of the video frame processing locally at the embedded device. Nevertheless, the performance evaluation of CNN-based models on embedded devices has been overlooked. In this paper, we provide a rigorous and practical framework for evaluating object detection models on embedded devices within a smart transportation system application. We conduct an extensive comparative study of six state-of-the-art object detection models on embedded devices, under different video frame conditions and devices' characteristics. Our evaluation framework can serve as a valuable resource for researchers and practitioners working in the fields of computer vision and embedded systems.