Research On Multiple Target Tracking Algorithm Based On Hash

With the vigorous development of the field of image processing,multiple object tracking as an important research direction of image processing have also made great progress.Although multiple object tracking technology has been successfully applied to various kinds of real-time video scene analysis,such as,unmanned drones,automatic driving.But at present in multiple target tracking algorithm still exist keep out occlusion,uncertain number of the target,data association and the requirements on real time behavior problems.In order to solve the problem of target quantity is not sure,need to use a good performance target detector to detect object on the video sequence,to obtain all the object's location and quantity of every frame image.So this paper uses the convolutional neural network to get the pedestrian classifier,and then combined with selective search algorithm to detect the pedestrian.The traditional target detection algorithm usually extracts the artificial features of the target first,then uses the feature to train a classifier,and finally uses the sliding window to obtain the candidate regions and classify them.But the traditional object detection algorithm has the following defects:One the one hand,the artificial feature extraction method is complex,and it is necessary for the designer to have some prior knowledge.One the other hand,the traditional target detection algorithm processes the feature extraction and object classification independently,if the extracted features of descriptive not enough,that classification algorithm also can't achieve better effect.But convolution neural network does not need to input the complex artificial feature,it can be directly input the sample image,by convolution operation autonomous learning to more natural,more generic features,Moreover,the feature has certain invariance for deformation,hence make convolution neural network is widely used in object detection.This article is based on the classical convolution LeNet-5 neural network model with the Caffe framework structures to construct convolution neural network,then through the common pedestrian detection data set select sample image to construct data sets,and carries on the contrast experiment on this data set,through the experimental results show that the convolution neural network can be applied to pedestrian detection as well.In order to solve the target occlusion problem,this paper will track.the target from there to leave the camera shooting range,the process is divided into initial,tracking,lost,the end of the four states,and in different states of different target to solve the occlusion problem.In order to solve the data association and practice requirements of the problem,the image features of the detected object of encoding using a hashing algorithm to get the hash code of the detection object,and then use the hash code inner product is used to measure the similarity between the current frame detection object and the previous time tracking target,select the largest combination of similarity to complete the target association.This not only simplifies the complexity of the algorithm,but also completes the association between them.Inorder to improve the accuracy of data association,based on the prior knowledge of tracking objects must maintain spatial continuity,the paper uses the centroid distance between the current frame detection object and the previous frame tracking target also as the evaluation criteria of similarity.Finally,experiments on MOT Benchmark datasets are carried out and compared with other multi-target tracking algorithms to verify the effectiveness of the proposed algorithm.At last,the research content of this paper is summarized,and based on the experimental results,this paper propose the improvement method of the multiple object tracking algorithm based on hash algorithm proposed in this paper.