Research On Multi-Target Tracking Algorithms In Intelligent Video Surveillance

Intelligent video surveillance systems use computer vision and pattern recognition technologies to process,analyze,and understand images in surveillance scenes,with huge social and economic benefits.Target tracking is a core technology in intelligent video surveillance systems,aiming for accurate target tracking based on target detection,of which the most valuable directions for research are single-camera multi-target tracking and multi-camera multi-target tracking.In single-camera multi-target tracking,occlusion often occurs between targets,and joint detection and re-identification is extremely time-consuming to train;while in cross-camera multi-target tracking,the performance is extremely dependent on the effect of single-camera multi-target tracking,and the differentiation of features of the same target in different viewpoints makes ID synchronization difficult.To address the above issues,the specific research in this paper is as follows:(1)A multi-target algorithm based on probabilistic expansion and improved OIM(Online Instance Matching)loss is proposed.To address the problems of detection and Re-ID branch feature misalignment and long training times for joint detection tracking(JDT)mode in single-camera multi-target tracking,this algorithm uses a deep aggregation network to combine multi-level features in the feature extraction phase to alleviate the feature misalignment problem and uses a Gaussian kernel function to probabilistically expand the samples in the target center range in the detection phase to shorten the training time.To address the problem of identity switching due to mutual target occlusion,this algorithm incorporates triples in the re-identification phase to augment the online instance matching loss of the network and utilizes motion and appearance features with Kalman filtering in the association phase to achieve efficient online association and improve tracking performance when targets are mutually occluded.The algorithm achieves MOTAs of 60.1%,70.7%,and 74.2% on the MOT15,MOT16,and MOT17 test sets,respectively,which are higher than the comparison algorithms.(2)A multi-camera multi-target tracking algorithm for synchronizing the same target ID is proposed.Multi-camera multi-target tracking often relies too much on the performance of single-camera multi-target tracking,and it is more difficult to synchronize the same target IDs under different camera views.To address the above problems,the high-precision single-stage detection method,namely Efficienr Det,is utilized to obtain the position information of the target and crops the image according to the detection results.It then uses the trained Re-ID model to extract the Re-ID appearance features of the target and uses the Io U and BIo U similarity of the detection frame and the Re-ID features through the improved Byte Track cascade matching strategy for data association to obtain robust multi-target tracking tracks in a single camera.Secondly,a set of triple vectors containing time frames,position information,and Re-ID appearance features is generated for each trajectory in single-camera video,and the similarity is calculated using the set of triple vectors in neighboring cameras to form a distance matrix to synchronize IDs of the same target trajectory in different cameras to generate multi-camera multi-target tracking trajectories.The higher-order tracking accuracy(HOTA)of this algorithm reaches 63.6,81.43,and 81.37% on Campus-04 and Campus-07 sequences of the MOT17 dataset and CVLAB EPFL dataset,respectively.In this paper,two multi-target tracking algorithms for intelligent video surveillance systems are proposed separately according to the camera settings,which can cope with different scene changes and thus achieve more accurate and robust target tracking.