Body Parts Detection for People Tracking Using Trees of Histogram of Oriented Gradient Descriptors

Vision algorithms face many challenging issues when itcomes to analyze human activities in video surveillance applications.For instance, occlusions makes the detectionand tracking of people a hard task to perform. Hence advancedand adapted solutions are required to analyze thecontent of video sequences. We here present a people detectionalgorithm based on a hierarchical tree of Histogramof Oriented Gradients referred to as HOG. The detectionis coupled with independently trained body part detectorsto enhance the detection performance and to reach state ofthe art performances. We adopt a person tracking schemewhich calculates HOG dissimilarities between detected personsthroughout a sequence. The algorithms are tested invideos with challenging situations such as occlusions. Falsealarms are further reduced by using 2D and 3D informationof moving objects segmented from a background referenceframe.