Research On Pedestrian Detection And Tracking Technology Based On Binocular Vision

In order to avoid collision with the pedestrian in man-machine co-integration environment,the obstacle avoidance system of the automated guided vehicle needs to accurately obtain the position and movement track of the pedestrian.In the complex working environment,the pedestrian detection methods based on vision are prone to false detection due to background interference.Combined with three-dimensional coordinate information of objects in the scene,a large amount of background information is removed,which overcomes the problem of false detection caused by complex background interference.When tracking the pedestrian target,the scale estimation method,the occlusion judgment method and Kalman filter algorithm are combined to overcome the problem of pedestrian loss caused by scale change and occlusion.The main work of this thesis is as follows:A pedestrian detection algorithm based on disparity map segmentation and feature optimization is proposed,which improves the problems of high false detection rate and long operation time of traditional pedestrian detection algorithms.The algorithm firstly uses the segmentation method based on interference region removal on the disparity map to obtain the region of interest,which reduces the area to be detected and avoids the error information beyond the detection distance range.Then,principal component analysis(PCA)algorithm is used to reduce the dimensionality of the histogram of oriented gradient(HOG)feature obtained from the region of interest,so as to avoid the interference caused by the redundant information in HOG feature on the performance of support vector machine(SVM)and accelerate the classification speed of SVM.Finally,the HOG-PCA feature and regionconstrained HSV color space feature are combined in series to improve the ability of the feature to characterize the pedestrian.The experimental results on the self-built INRIASTEREO dataset show that compared with the traditional HOG+SVM algorithm,the detection efficiency and accuracy of the proposed algorithm are significantly improved,and the precision and recall are 98.76% and 98.89% respectively.The proposed algorithm effectively overcomes the problem of false detection caused by long-distance complex background interference,and also improves weak detection effect caused by the changeable pedestrian posture to a certain extent.A scale-adaptive and anti-occlusion kernel correlation filter(KCF)algorithm is given,which overcomes the problem that the KCF algorithm is prone to tracking failure when scale of the pedestrian changes or the pedestrian is occluded.In the algorithm,three scale factors are used to estimate the scale of the pedestrian,which realizes the scale-adaptive function;the occlusion judgment method based on F-APCE is used to detect the occlusion degree of the pedestrian.When the pedestrian is partially occluded,the model is not updated,and when the pedestrian is seriously occluded,the Kalman filter is used to predict the position of the pedestrian.The experimental results on OTB100 dataset show that compared with KCF algorithm,the accuracy and success rate of the improved algorithm are increased by 6.75%and 15.88% respectively,and the improved algorithm can still accurately track the pedestrian in scale change and occlusion scenes.