Detecting humans in videos by combining heterogeneous detectors

Human detection is an actively researched problem in computer vision, and it has driven the development of a number of detectors, both problem-specific and general purpose. We investigate the performance of human detection in video sequences when three commonly used detectors are combined: background subtraction, boosted cascades of Haar-like features and histograms of oriented gradients. We qualitatively analyze the performance of each of the detectors and quantitatively estimate the performance gain obtained by combining all three. We employ sequences from the publicly available TUD dataset. HOG and background subtraction perform reasonably well when used individually, while boosted cascades of Haar-like features performs inadequately in terms of localization accuracy and false positive rate. Our results indicate that overall detection precision and recall can be improved when combining HOG with background subtraction. Adding the boosted cascades of Haar-like features does not contribute to the overall performance. We conclude that combining HOG and background subtraction is an elegant solution for utilizing the advantages of video data to improve detection performance while maintaining a low processing cost.