Pedestrian detection in digital videos using committee of motion feature extractors

Human detection in digital images is a challenge because the motion of the camera, background and variations in pose, appearance, clothing and illumination introduce difficulties for person detection. Several pedestrian detectors were proposed recently, such as the Aggregated Channel Features (ACF). These types of detectors are based on features related to the shape of the object. These detectors generate many false alarms. In this paper we propose the use of motion features in the ACF framework to mitigate false alarms emitted. Three motion features are proposed: WSTD, MBH and IMHcd. We combined these features with ACF. Then, committees of classifiers were created from these feature combinations improving original ACF results and reducing false positives per image (FPPI). An improvement on the miss rate for 100 FPPI and on the log-average miss rate was obtained, reducing these values in 19 and 8.46 percentage points respectively on Caltech dataset.