Research And Design Of Moving Target Recognition And Tracking Based On Machine Learning

Motion object recognition and tracking based on machine learning is one of the hot research topics in the field of machine vision and pattern recognition.As a kind of system with early warning,prevention and active monitoring functions,the target recognition and tracking system based on UAV video can effectively solve the practical problems such as not timely and misjudgment in the manual processing.This paper closely follows the core issues of moving target recognition and tracking,and studies the three aspects of target recognition,target tracking and recognition and tracking system design.Firstly,aiming at the problem of low recognition accuracy,slow rate and missed detection in target recognition process,an improved target recognition algorithm based on random forest and support vector machine is studied.The algorithm affects the accuracy of recognition for single feature.The color feature(Lab)is merged with the improved directional gradient histogram(FHOG)feature,and then the target region of interest is derived from the random forest,and then identified by the LIBLINEAR classifier of the linear support vector machine.The experimental results show that the recognizer in this paper has a 9.35% higher recall rate than the traditional support vector machine in pedestrian recognition,and the missed detection rate is reduced by 0.68%,and the rate reaches 2.70 frames/second.Secondly,in order to improve the tracking performance of the correlation filtering algorithm(ECO-HC)which combines the gradient histogram features and color features,a correlation filtering algorithm based on feature fusion and adaptive learning rate improvement is studied.The algorithm determines the feature fusion weight based on the characteristics of the gradient histogram features and color features and its influence on the tracking performance of the UAV's pedestrian tracking.At the same time,the adaptive learning rate method is adopted to enable the tracker to adaptively cope with complex target motion problem.The experimental results show that the improved algorithm has 3.3% and 2.6% higher average distance accuracy and overlap rate than ECO-HC,and the tracking speed reaches 51.8 frames/second.Finally,this paper designs a set of target recognition and tracking system,which uses the recognizer to identify the target in the first frame image,so that it can be used as the input of the tracker for subsequent tracking.For the situation that new target may appear during the tracking process.The system adopts the method of interval frame start identifier,multi-target parallel tracking by data association algorithm,and the Kalman filter algorithm for auxiliary supervisory tracking to ensure tracking accuracy.The system is verified by self-built UAV video set,and the results show that the system has robust performance and real-time performance.