An Optimized Hardware/Software Co-Design Framework for Real-Time Pedestrian Detection

Abstract

Real-time Pedestrian Detection is an important feature of on-board intelligent Advanced Driver Assistance System (ADAS). The system employs feature detection and classification to identify human objects for pedestrian detection and their collision avoidance. However, these processes are computationally expensive and too complex for real-time implementation on a general-purpose processor. To address the issue of real-time deployment of such systems, this paper presents an approach using hardware/software co-design based on Xilinx Zynq FPGA and optimized OpenCV implementation. It uses Histogram of Oriented Gradients (HOG) features for human object detection and multi-stage Support Vector Machine (SVM) for object classification. The results show that our proposed co-design approach is at least 7× lower in latency than the reported software-based implementations and 6.6× faster than the previously reported HW/SW co-design approaches. In addition, our solution achieves 94% recall and 92% precision for $640\times 480$ resolution images. The results have been verified using INRIA and MIT datasets.