HOG pedestrian detection applied to scenes with heavy occlusion

This paper describes the implementation of a pedestrian detection system which is based on the Histogram of Oriented Gradients (HOG) principle and which tries to improve the overall detection performance by combining several part based detectors in a simple voting scheme. The HOG feature based part detectors are specifically trained for head, head-left, head-right, and left/right sides of people, assuming that these parts should be recognized even in very crowded environments like busy public transportation platforms. The part detectors are trained on the INRIA people image database using a polynomial Support Vector Machine. Experiments are undertaken with completely different test samples which have been extracted from two imaging campaigns in an outdoor setup and in an underground station. Our results demonstrate that the performance of pedestrian detection degrades drastically in very crowded scenes, but that through the combination of part detectors a gain in robustness and detection rate can be achieved at least for classifier settings which yield very low false positive rates.