Research And Application Of Multi-target Tracking Algorithm Based On Multi-feature Fusion

Multiple objective tracking has always been one of the most important problems in computer vision research,which provides support for recognition tasks.At present,many mult i objects tracking algorithm focuses on pedestrians and vehicles.In this paper,a multi-face tracking algorithm based on multi-feature fusion is proposed for the security surveillance video at Nanjing Metro Station.The main work of this paper is as follows:1.Based on the MTCNN algorithm,this paper proposes a fast and accurate face detection method.In order to solve the problem of slow convergence during training in the origina l network structure,this method uses Batch Norm structure in the network structure to normalize the data in the training process.At the same time,this method deletes the feature regression points in the MTCNN network,and improves the computational speed in the forward network.Because the original MTCNN network uses mean squared error loss function,the network converges slowly due to some points which are far away from detected ones.This method uses first-order loss function to avoid this problem.In network training,distilling knowledge-based network and cascaded detection method are used to further compress the size and improve performance of network.Extensive experimental evaluation on the FDDB datasets show that,the result of this method has been improved by 8% compared with the original MTCNN network,with a fast speed of 30 FPS in real time.2.Based on the improved MTCNN face detection method,this paper proposes a mult itarget tracking algorithm based on multi-feature fusion.Compared with traditional handcraft features,in order to better cope with the chal enges of target scale change,target occlusion,deformation and rotation,the convolution features in MTCNN network are taken as appearance features,the state feature and position scale feature obtained by Kalman filter are used as the fusion features for data association.At the same time,considering the different cases of tracking failure,a data association method based on conditional judgment is proposed,which can track multiple targets in three situations: tracking success,tracking failure and object overlapping.Experiments show that this method has great improvement compared with the multi-tar get tracking algorithm using single feature and general data association method.3.A multi-target tracking security monitoring system based on C++ is designed.The system mainly includes dynamic retrieval module,query module,resource manageme nt module and parameter management module.It completes the functions of query and retrieval,tracking and statistics for Metro security monitoring.At the same time,the system has two architectures,C/S and B/S,which can solve various problems for different users.