Embedded Real-Time Vehicle and Pedestrian Detection Using a Compressed Tiny YOLO v3 Architecture

Vehicle and pedestrian detection (VaPD) is one of the most critical tasks in an advanced driver assistance system which help the driver to drive safely and save the pedestrian life. VaPD is a typical object detection problem that requires a trade-off among accuracy, speed, and memory consumption. Most existing methods focus on improving detection accuracy, while ignoring VaPD requires real-time detection speed with limited computational resources. Thus, it is of primary importance to study light-weight and real-time VaPD methods for embedded devices, that is hardware platforms with limited computation and memory resources. To deal with these issues, this paper proposes a low-rank (LR) Tiny YOLO v3 architecture that meets the requirements of real-time VaPD on embedded systems. The architecture has been developed starting from Tiny YOLO v3 adopting a convolutional neural network compression technique based on Tucker tensor decomposition, able to reduce the computational complexity of the network. A wide experimentation has been carried out on two embedded platforms, Raspberry Pi 4 and NVIDIA Jetson Nano 2 GB, and two datasets commonly used for VaPD, PASCAL VOC and KITTI dataset, showing the superiority of the LR Tiny YOLO v3 with respect to the state-of-the-art networks in obtaining the best compromise between inference time, accuracy and memory occupancy. Moreover, the proposed architecture meets the requirements of VaPD on embedded systems using only 22% of the memory required by the baseline Tiny YOLO v3 Darknet, and always providing better inference time (36.46 FPS) with only a marginal decrease in accuracy (~2%).