Vision-based Online Multiple Pedestrian Tracking In Single Scene

The visual monitoring system has been popularized in a few years,but it is still in the low development level of artificial retrieval and comparison analysis.Especially,due to the change of illumination and appearance caused by the limitation of vision in single camera,the inter-leaving of dynamic targets and frequent occlusion will cause great influence on the performance of tracking system.Therefore,the multi-objective monitoring system based on vision has great academic value and application prospect.In this thesis,the research on single scene and multi-target reliable association is carried out,and the solution of multi-camera and multi-target tracking is put forward.The reasonable construction of pedestrian appearance and motion model for the scene needs is very important for the recognition and tracking of pedestrians.However,the traditional appearance model does neither construct the spatial information and appearance together,nor does it analyze the change of the pedestrian occlusion.Based on the traditional features,the further introduced local features and the information of multiple spatial structure features extraction in different positions,moreover,this thesis carries on the research and improvement of the above problems.For tackle short-term occlusion orlong-term occlusion,this thesis divided pedestrian features into general appearance feature and the detail one.For the moving target model,the current research does not take into account the fluctuation caused by the pedestrian steps.In this thesis,the traditional Kalman prediction model is decelerated,which ensures the stability of tracking.At the same time,the motion fluctuation is reduced by means of mean and reasonable timing information.It is necessary to identify all pedestrian targets using reasonable motion prediction and accurate pedestrian appearance.Generally,it can be matched by distance similarity function.In this paper,the two point discriminant analysisalgorithm(XQDA)based on cross view is adapted,and all similarity is further trained and matched.The recognition degree between pedestrians is increased 3?6 times.Then,the general Hungarian algorithm is used to match,and the higher pedestrian matching accuracy and tracking accuracy are obtained after testing.For tracking framework,based on the general framework,we improve the three parts of target initialization,reducing matching range and tracker classification,making pedestrian tracking more stable and logical,and improving tracking performance.The experimental results show that,although the initialization process of the algorithm is not tracked at the beginning stage,it can effectively reduce the error detection.At the same time,the precision of matching algorithm is improved by reducing the matching range,at the same time,the computational complexity is reduced,thus the operation speed of the target tracking algorithm is improved.Moreover,the tracker classification ensure that pedestrians leave the scene to match again,or pedestrians meet the occlusion and so on.The experimental results show that the tracking framework proposed in this paper is more stable,and the tracking accuracy can be improved by 5?10%compared with the original algorithm.Finally,this paper summarizes the research content and work,and looks forward to the main research direction in the future.