Research On Real-time Pedestrian Detection Algorithm In Vehicle Assisted Driving System

As an important direction of automobile development,the Advanced Driver Assistance Systems(ADAS)has undergone more than ten years of technological development,and its functions are constantly enriched and its reliability is constantly improving.Pedestrian detection based on computer vision has become one of the most active research topics in the field of computer vision and intelligent vehicles,due to its important application value in vehicle assisted driving systems.In this paper,the related algorithms of the pedestrian detection module in the vehicle assisted driving system are deeply studied,and a pedestrian detection system that can be deployed is proposed.In practical applications,pedestrian detection faces more difficult challenges.Mainly reflected in the following aspects: Firstly,the scene is complex,including obvious target differences,motion blur,small target detection and dramatic changes in lighting,etc.,so requires the robustness of the algorithm is high;secondly,it is limited by the cost of the vehicle,auxiliary It is difficult for the driving system to use Lidar to perform rapid target initial determination like an unmanned vehicle,which lacks a good method for determining the region of interest.Thirdly,the system is a real-time working system,which requires a limit of time cost and complexity of the algorithm.if the algorithm or the model used is too large,and the real-time performance is difficult to meet;Finally,using high-performance compute equipment to deal with,in the vehicle installation process,the volume is too large,the heat dissipation performance is poor,and the cost is greatly increased.In order to solve the above problems,this paper proposes a real-time pedestrian detection framework.The solution mainly includes the following parts:(1)Determination of ROI:Based on the groundHog algorithm,this paper adopts a fast multi-scale adaptive pedestrian ROI extraction method,and gives the engineering implementation code.The acceleration effect of the framework is very obvious;(2)Detector: This paper proposes a cascade approach to alleviate the contradiction between speed and accuracy.In the process of DPM improvement,a supervised threshold selection method is proposed,which has better robustness than the previous unsupervised threshold selection method.Finally,the experiment proves that it is acceptable to remove unnecessary model’s loss to the performance of the algorithm,and the improvement of speed is obvious;(3)Multi-target tracker: In order to better judge the drift and loss of the target,a pre-trained pedestrian classifier is proposed,which takes into account the survival time and the results of the classifier.Finally,the drift or loss of the target is judged,which is more reliable than the previous response threshold or peak-to-side lobe ratio method.At the same time,in order to adapt to the scale changes of pedestrians in real scenes,a coarse-to-fine weight-biased geo-scale search is proposed to achieve more accurate approximation of the real target scale;(4)In order to solve the problem of system runaway caused by speed mismatch between detector and tracker when frame acquisition is triggered by hardware,a polling access method for image subarea is proposed in the end;(5)A pedestrian detection system with excellent performance is constructed in this paper.