Object Tracking In Crowded Scenes

Object tracking in videos serves as a fundamental and important task in computer vision.It has profound significance of research in academe and broad prospect of ap-plications in industry.In this thesis,we make some research about the object tracking in crowed scene based on the tracking-by-detection framework.Focusing on the missing detections,data association and object update,we make intensive research,and achieve results as following:(1)We fuse the tracking results from the Discriminative Correlation Filter Tracker with Channel and Spatial Reliability to track multiple objects.We find that a big dis-advantage of tracking-by-detections methods is that the detector can not locate objects when they are occluded or too small.The missing detections can fragment the complete trajectory,and increase the probability of target identity switching,then cause tracking results bad finally.Therefore,we propose to fuse the correlation filter tracking results with the detections and provide the predictions for all targets.Even the detector can not locate the targets,they also can be found by the correlation filter tracker.In addition,we also apply the Kalman filter to perform coordinate fusion of single target tracking results and matched detections.This can reduce effects of noise,and smooth motion trajectory of targets.(2)We apply Markov Clustering to associate data.For the case that more than one detections matched the target can be initialized as a new target,we borrow the idea from local clustering.We apply the targets and detections to the Markov model,and use Markov Clustering algorithm to get the clusters which include matched target and detections.Each cluster has one target and many detections,making extra detections can match with one target.So it can reduce the number of false positives,and improve tracking performance.Furthermore,we change the convergence condition of clustering,and design an accelerated Markov Clustering algorithm,so that to shorten the cost of algorithm running time.(3)We employ the resnet on feature pyramid networks to implement the bounding box refinement.The target location is updated by the associated detection,so the de-tection influence the precision of target location directly.Since to improve the precision of target trajectory,we refine and adjust the detection bounding boxes by convolutional neural network.We apply the resnet on feature pyramid networks to extract the deep feature of the video frame,and make detections as region proposals,then get each feature map of detections by region of interest pooling.According to the pooled feature map,the network regresses the refined bounding boxes which make the final target location more precise.