Research On Pedestrian Tracking Algorithm Based On Machine Vision

With the development of intelligent connected cars,assisted driving systems have become the focus of research.Pedestrian tracking algorithm is an important part of assisted driving system.It uses kernel correlation filtering algorithm and convolutional neural network related technology.In this paper,multi-feature fusion anti-occlusion pedestrian re-identification tracking algorithm and deep convolution feature fusion scale adaptive pedestrian tracking Methods,focusing on the accuracy nature of pedestrian tracking,the main research contents are as follows:(1)Create a new mixed data set OTB100-UP&T.By collecting samples of pedestrian images that meet the driving perspective on multiple roads in the university campus and Xi'an urban area,additional samples are solved to solve the defects of the original data set OTB100 with a small number of samples and single data.(2)The multi-feature fusion anti-occlusion pedestrian re-identification tracking algorithm is adopted to solve the problem that the algorithm is not easy to continue to track pedestrians under the conditions of pedestrian occlusion and fast pedestrian movement.The direction gradient histogram(HOG)and scale-invariant feature transform(SIFT)features are introduced into the kernel correlation filtering(KCF)tracking algorithm.When the target pedestrian in the video frame is partially occluded,the SIFT and HOG features of the current frame are extracted.The principal component analysis(PCA)dimensionality reduction technique is used to reduce the dimension of the extracted HOG features and SIFT features,and the HOG-SIFT feature template is synthesized by the cascaded dimension reduction of HOG features and SIFT features.Use the formed feature template to match the HOG-SIFT feature in the next frame to quickly relocate and track the pedestrian.(3)Scale adaptive pedestrian tracking method using deep convolution feature fusion.Solve the problem that the tracker cannot adaptively track when the pedestrian in the relevant filtered pedestrian tracking algorithm rotates,deforms or scales,and the tracking accuracy is low.Combining convolutional features with correlation filtering,extracting multi-layer convolutional features of pedestrian candidate regions through deep convolutional network architecture,and constructing a two-dimensional positioning filter using kernel correlation filtering algorithm to obtain multi-layer convolutional features and weighting Fusion,determine the location of pedestrians,and then extract the multi-scale pedestrian direction gradient histogram features,estimate the best pedestrian scale,and achieve fast,accurate and continuous pedestrian tracking.The experimental results show that the multi-feature fusion anti-occlusion pedestrian re-identification tracking algorithm in the case of occlusion and motion blur,the algorithm's tracking distance accuracy reaches 72.4%,and the success rate reaches 58.7%.The scale-adapted pedestrian tracking algorithm with deep convolution feature fusion has a precision of 88.2%in tracking distance and a success rate of 61.3%in the case of pedestrian rotation and scale change,which has reference significance for the research of assisted driving systems.