Research On OpenCL Fast Pedestrian Detection Method Based On CENTRIST Feature

Pedestrian detection had been a hot research field over ten years due to its great potential in several important applications such as driving safety,surveillance and robotics.Driving safety becomes a urgent need along with the explosion of native car market,ADAS implemented by high technology is one of the important methods,this paper would focus on pedestrian collision avoidance in ADAS.Pedestrian collision avoidance relies on robust and real time pedestrian detection method which is the most hard in pedestrian detection problem.Classic detection method use HOG features which is difficult to deal with real time task because of its complexity of computation,recently CENTRIST features is proposed for pedestrian detection task,CENTRIST means Census transform histogram,is designed to describe critical contour information,overcame the inefficiency of HOG feature and is proved to be much more suitable for pedestrian detection task than HOG.C4,an implementation of pedestrian detection based on CENTRIST feature,using fast linear SVM classifier and fast HIK SVM classifier achieved both higher accuracy and much more efficiency.CPU in embed system always had low performance than PC,the sequential C4 algorithm is hard to meet the real time requirement,this paper analyzed popular methods to accelerate image process,proposed an parallel C4 algorithm implemented by OpenCL,the parallel detection method make full use of current hardware,achieved a high efficient detection method in embed system at low cost.This paper dived deep into heterogeneous computing solutions for image processing,both hardware and software are mentioned here,the analysis showed the possibility to achieve real-time human detection based on OpenCL parallel C4 algorithm ran even on an embedded board.Parallel C4 algorithm based on heterogeneous computation converted the original image to grayscale at first,an resize the grayscale by fixed ratio for several times to build the image pyramid,then detect every image in the pyramid and get all detection windows.In detection phase,this paper proposed parallel Sobel operator,CT feature extractor,auxiliary image computation,integral image algorithm,linear SVM classifier and HIK SVM classifier all based on OpenCL.Experiments showed parallel detection algorithm achieved the same detection accuracy as the sequential detection method,but parallel detection method ran more than 10 times faster on PC and nearly 50 times faster on embedded board.Parallel C4 detection method based on OpenCL ensures the portability of functionality,when this algorithm was applied in other hardware platform,only a little customize for different hardware features was required to achieve performance refinement,so it provide great convenience for cross platform solution..