Research On Pedestrian Detection And Tracking Algorithm Based On Head And Shoulder Model

With the development of computer vision,artificial intelligence and other related technologies,pedestrian detection and tracking technology has received more and more attention.However,because of the changeable characteristics of the pedestrians,the appearance be affected by factors such as clothing,scale changes,occlusion and shooting angles.Traditional single feature or model is difficult to achieve accurate pedestrian detection,but the shape of the pedestrian's head and shoulders is relatively stable.This paper proposes a head-and-shoulder classifier to perform the pedestrian detection.Targeting the fast and efficient KCF tracking algorithm cannot be adaptive to estimate the target scale,the combination of algorithm improvement and software design to achieve the scale of adaptive multi-target tracking.The content of this article mainly has the following parts:(1)Firstly,proposed the system framework.Focus on the key technologies in pedestrian detection and tracking,a variety of pedestrian detection and tracking algorithms are analyzed and researched.Finally,a head-and-shoulder classifier with a combined structure proposed for pedestrian detection,and an improved KCF tracking algorithm is used.(2)Design a head and shoulders classifier of the combined structure.In view of the changeable appearance of pedestrians and the fact that their head and shoulders can remain relatively stable,a head and shoulder classifier using a combined structure proposed for pedestrian detection.Using the Adaboost algorithm train the initial detector with Haar-like features to obtain the candidate regions.The second detector trained by support vector machine(SVM)algorithm with the Histogram of Oriented Gradients(HOG)feature to accurately obtain head and shoulder position information.Through experiments,verified that the combined structure classifier is superior to traditional one-stage detection in detection accuracy and false detection rate.(3)Improved KCF tracking algorithm.This article focuses on real-time requirements.Therefore,the tracking algorithm uses Kernelized Correlation Filters.At the same time,aiming at the problem that the algorithm cannot adapt to the target and can only perform single target tracking,an improved KCF tracking algorithm proposed.By adding three scale factors,the target's scale estimation was achieved.Comparing with the original KCF algorithm and other tracking algorithms,this method has better robustness and tracking accuracy when the target scale changes.(4)Realization of the Pedestrian Detection and Tracking System.The use of C++ language combined with VS2013 completed the development of the system.Statistics on the number of people in the target area further validated the effectiveness of the system.