In this paper, we introduce a fully automatic system to detect and track multiple pedestrians in highdensity crowds in the occurrence of severe occlusion. Typical approaches such as frame differencing and body partbased pedestrian detection is unsuccessful when most of the scene is in motion and body parts of the pedestrians are occluded. To overcome this difficulty, we incorporate human detection and tracking into a single framework using modified Gaussian Mixture Model (GMM) based Particle Filter tracking and introduce a confirmation-byclassification method for tracking associates detections with tracks, tracks humans through occlusions, and eliminates false positive tracks. For detection, instead of traditional GMM, we use modified GMM algorithm in which parameters are tuned automatically and appropriately according to type of background in video sequence. Particle Filter is applied to handle the dense crowd tracking accurately and color histograms is used for appearance modeling. To further reduce false detections due to dense features and shadows, we introduce cross covariance method for estimation and utilization of a state of objects that reduces false positives while preserving high detection rates. In an experimental evaluation, we show that our proposed system enables the construction of an excellent pedestrian tracker for dense crowd.
Volume 11 | 06-Special Issue
Pages: 1522-1535