Research And Implementation Of Pedestrian Detection System Based On APEX

Pedestrian detection based on computer vision has become one of the most active research topics of intelligent vehicles’ field, which has more and more important application value in Driver Assistance Systems. The purpose of vehicle-mounted pedestrian detection system is to capture frame using installed cameras in the moving vehicles, to estimate the potential risk using pedestrian detection algorithms, to alert for possible dangerous situation beforehand and to take strategy to protect pedestrians.This thesis states research and implementation of a pedestrian detection system based on APEX. It’s a subsystem of Advanced Driver Assistance System (ADAS), which is a actual project developed by Advanced Automotive Electronic Technology Research Center(AAC) of Neusoft. The system is implemented in APEX platform provided by Freescale, which is a famous micro controller company in Germany.This thesis describes a cascade Adaboost classifier for pedestrian detection. In the offline training stage, Haar-like features are used for training Adaboost classifier; in the online detection stage, a trained cascade Adaboost classifier is used for classify muti-scale input images. Then merge the results from different scaling images to get the final detection results.I draw a conclusion that integral image algorithm, rotated 45 degree integral image algorithm and Adaboost classification algorithm can work and transplant to APEX in parallel with some modification after deeply analyze Haar and Adaboost algorithms, which belong to pedestrian detection algorithms. Then, the purpose is realized.The innovation of this thesis is achieving parallelization and modularization for pedestrian detection algorithms in APEX, which has 32 CUs working in parallel at the same time. By testing, the system process a image with size of 640x392 gray scale to judge whether or not containing pedestrians takes about 30ms at 400MHZ clock frequency of APEX. Therefore, the system improves performance by about 8 times and acquires better real-time performance and robustness comparing with realization in ARM platform.At the last chapter, the system tested in different scenarios, the results of evaluation show it can achieve good detection results. Since its small and easy to control, the system has a wide application prospect.