The Research And Application Of Moving Multi Target Detection And Tracking Based On Deep Neural Network

Video tracking is a new research direction in recent years.It integrates technology of computer vision,pattern recognition,AI and other disciplines,and has broad application prospects in security monitoring,intelligent transportation,video compression and retrieval.At the same time,with the learning and application of deep neural network,there is a new combination of object selection in video tracking and self learning in target tracking.Therefore,it is of great practical significance to apply the deep neural network to the detection and tracking of multiple targets.Multi target tracking is on of the research hotspots in the firld of video tracking,it can effectively solve the behavior judgment and crowding degree identification of multi-objective group,an effective way to improve the efficiency of the existing video target tracking,and ensure the accuracy of the existing single target tracking and accuracy.This thesis applied the depth of the neural network through pre training log a target image,the general goal can be simple to complex structural features of the said.By using the structural characteristics obtained from training,we can get classifiers that can track target by online training,and classify and select the target to track.Secondly,after selecting the target box,the tracker uses optical flow to estimate target motion according to the limited motion between target and frame.Location detector for each frame of image scanning to identify the current time has been observed and studied with the aim of similar appearance.For the false positive and negative samples that may appear in the detector,the learning module is used for self learning and the tracker,detector and learning module are constantly improved.According to the result of tracker,two kinds of errors of the detector are evaluated,and training samples are updated according to the evaluation results,and the key feature points of the tracker are updated at the same time.A set of algorithms for multi-target tracking detection in video in a variety of specific scenes is formed.Finally,The idea of combining MDP(Markov Decision Processes)and multi centroid multi-objective prediction algorithm,we establish a separate centroid decision making learning method for each target according to the "separation and combination" mode.The distance and direction of the target displacement are accurate by the frequency domain vector and the space vector,and the target action judgment mechanism is formed through the decision learning.The displacement and action of each centroid become the judgment and reference standard for the displacement and action of the target.The effect test is carried out under a variety of specific scene video and the application of multi target tracking detection in video within a variety of specific scenes is completed.