Implementation And Optimization Of Hi3531-based Fast Pedestrian Detection

With the rapid development of economy and society,the city's public security system is becoming more and more important.Video surveillance technology is gradually moving towards intelligent and plays a more and more important role in the public security system.Pedestrian is the main target in video surveillance in which case the research of pedestrian detection technology is of great significance.This paper first introduces the overall architecture of Hi3531-based pedestrian detection system and the general situation of software and hardware development platform.It introduces the general process and the basic principle of the pedestrian detection algorithm including the input of images and detector,the construction of pyramid,screen of the regions,etc.Based on the fast pedestrian detection algorithm,this paper introduces the transplantation into the Hi3531 platform including the estimation of gradient histogram in multi-scales and the implementation method for fast pedestrian detection.The feasibility of the fast pedestrian detection algorithm is demonstrated from the theoretical point of view.In this paper,the establishment of Hi3531-based pedestrian detection module is introduced in detail.Firstly,This paper introduces the configuration of the development environment and how data is transferred in the system.The results of Hi3531-based pedestrian detection are given.Multi-stream transmission and related storage mechanism for semantic retrieval application are introduced in detail.Based on different bit stream and different detection results,the video stream and the description of the flow do the corresponding processing.In this way,this paper gives a more detailed introduction of the implementation methods.Based on the HiMPP media processing software development reference,this paper presents optimization scheme for the pedestrian detection utilizing Hi3531 hardware platform and the basic operators provided by intelligent video engine speedup module.In order to reduce the algorithm time for color space conversion,this paper introduces developing a module capable of identifying the features of different color spaces to facilitate pedestrian detection.Experimental results reveal that the detection time,running on Hi3531 platform,is remarkably reduced after a series of optimization that induces no offset to detection accuracy though.