Research On Detection Based Online Multi-object Tracking

With the rapid development of science and technology and the improvement of people's living standard,the awareness of security prevention is gradually popular.In the field of security prevention,video surveillance plays an important role.In recent years,the network of video surveillance gradually spread to every corner of public places and also began to enter into thousands of families.At present,the widely used video surveillance system is semi-manual,which is low degree of automation and needs a lot of human and material resources.Intelligent video surveillance system is bound to become a future development trend.Multi-object tracking technology is a hot issue in computer vision,and is also one of the most key technologies in intelligent video surveillance.Multi-object tracking aims to provide the location and trajectory of interest targets in video sequences and is the basis of follow-up target recognition,target behavior analysis and understanding.With the improvement of target detection technology,multi-object tracking based on detection responses has become the research trend,and has achieved some good results.However,due to the complex background of video surveillance and frequent occlusion,some difficult problems are still not solved.This dissertation focuses on some of these problems and puts forward an improved online multi-object tracking method based on detection,the main research content and innovation points are as follows:1.We proposed a hierarchical data association algorithm based on rapid location prediction.The online multi-target tracking algorithm based on detection is data association frame by frame essentially.These approaches demand for good detection responses and the missing responses have great impact on the tracking performance especially.In this dissertation,on the basis of the traditional online tracking algorithm,we established an individual predictor based on kernelized correlation algorithm for each target,and then predictor is associated with the detection responses to determine the state of the target in each frame together.In order to overcome the defects of the fixed scale of the kernelized correlation algorithm,the scale and the template is adaptively updated based the corresponding detection response information.The experimental results show that the algorithm proposed in this dissertation can overcome the influence of missing detection.2.We proposed an occlusion handling method based on particle swarm optimization algorithm with binding force.Occlusion handling is a difficult task in multi-object tracking.Because motion state of targets is always random,it is difficult to calculate the target's position accurately in occlusions which often results in tracking drift.However,if the occluded target has not been present,we can think it still stayed in the occlusion area.According to this priori information,we can predict occluded targets with the correct tracking targets.In this dissertation,particle swarm optimization algorithm is used to search the best position of occluded target.And we combined a binding force into iterative calculation of particle swarm optimization algorithm,so that the particle swarm is constrained near the occluded area,and does not happen to drift.The results show that the algorithm can overcome the tracking drift in occlusion to some extent.