A Lightweight Pedestrian Detection Model for Edge Computing Systems

Most vision-based pedestrian detection systems adopted deep learning approaches with the framework of convolutional neural networks (CNNs) to reach the state-of-the-art detection accuracy. As CNN-based approaches are computationally intensive, the deployment of those systems to a resource-limited edge device is a challenging task, especially for a real-time application such as intelligent vehicles. In this study, we proposed a lightweight high-performance edge computing solution to achieve rapid and accurate performance in pedestrian detection. Experimental results showed that the proposed framework can effectively reduce the miss rate of the YOLO-tiny detection model from 48.8% to 26.2% while achieving an inference speed of 31 frames per second (FPS) tested on the Caltech pedestrian dataset.