Research On Pedestrian Detection And Tracking Based On Binocular Vision

Nowadays,with the introduction of the concept of unmanned driving,there has been a worldwide research for artificial intelligence and machine vision triggered by intelligent driving.The driving safety problem of smart car is the key point in the research.The assistant driving system of the smart car integrates many technologies such as machine vision technology and artificial intelligence,in order to ensure the driving safety of the vehicle.Pedestrians belong to the most special category of obstacles.Therefore,in the course of realizing intelligent driving,the pedestrian's three-dimensional coordinate information needs to be acquired in real time to ensure effective and reliable pedestrian detection and tracking so as to reduce or avoid the harm to pedestrians.The traditional pedestrian detection based on color image sequences is seriously affected by noise such as ambient light,and the classification method using machine learning has the disadvantages of long operation time and high complexity.In response to this problem,this subject uses binocular vision technology to study pedestrian detection of positioning,classification and other issues.An improved cascade classifier model is proposed to realize the classification of obstacles.Using the correspondence between obstacle images in binocular disparity map and the actual obstacle targets in the world coordinate system,the actual obstacle detection is transformed into the straight line detection in the UV parallax map,combining the key points of clustering analysis and straight line fitting Technology to determine the region of interest(ROI)of the obstacle in the disparity map;and according to the determined ROI region,the image segmentation algorithm is used to extract the obstacle profile.The research shows that the histogram of the pedestrian contour projection has the same Gaussian distribution in the horizontal direction as the vertical direction.Based on this,a classifier with improved cascade is designed and constructed.The first level is the traditional vertical projection classifier,and the second level is the horizontal projection classifier designed in this project.Through the improved cascade classifier,the final detection rate of 92% is achieved,which improves 9% before the improvement.In order to solve the problem of traditional background tracking based on color image,such as tracking target is easily influenced by background,color image contains complicated information and processing time is long.In this paper,binocular disparity map is used as input of tracking system,track.The real-time and tracking errors of two types of pedestrian tracking based on binocular disparity map and color image are compared and analyzed by experimental simulation.The results show that the pedestrian tracking error based on binocular disparity is 0.71% higher than the color image tracking error,but the tracking real-time performance is improved 35.5%.