Design of Object Detection Systems on a SoC-FPGA Platform

Abstract

Object detection is a crucial and challenging task in the field of embedded vision and robotics. Over the last 15 years, various object detection algorithms/systems (e.g., face detection, traffic sign detection) are proposed by different researchers and companies. Most of these research are either focused on improving the detection performance or devoted to boosting the detection speed, which reveals two typically employed criteria while evaluating an object detection algorithm/system: detection accuracy and execution speed. Considering these two factors and the application of object detection to the domains such as service robots and Advanced Driving Assistance System (ADAS), FPGA is a promising platform to achieve accurate object detection in real-time. Therefore, FPGA-based robust object detection systems are designed in this work. The main work of this thesis can be divided into two parts: promising algorithm obtaining and hardware design on SoC-FPGA. Firstly, representative object detection algorithms are selected, implemented, and evaluated. Thereafter, a generalized object detection framework is created. With this framework, pedestrian detection, traffic sign detection, and head detection algorithms are realized and tested. The experiments verify that promising detection results can be obtained by employing the generalized object detection framework. For the work of hardware design on FPGA, the platform of the object detection system, which consists of stereo OV7670 cameras, Xilinx Zedboard, and a monitor that can visualize the detection results, is created. After that, IP cores that correspond to each block of the framework are designed. Configurable parameters are provided by each IP core so that the IPs, especially feature calculation IP and feature scaler IP, can be correctly instanced according to the fast feature pyramid theory. Finally, by employing the designed IP cores, pedestrian detection system, traffic sign detection system, and head detection system are designed and evaluated. The on-board testing results show that real-time object (e.g., pedestrian, traffic sign, head) detection with promising accuracy can all be achieved. In addition, with the generalized object detection framework and the designed IP-toolbox, the object detection system that targets any instance of objects can be designed and implemented rapidly.