Improving channel features using statistical analysis for pedestrian detection

As one of the most successful features in vision-based pedestrian detection, filtered channel features have drawn considerable attention in the research community. In this paper, we improve the channel features by performing a statistical analysis of each channel and taking both the average map and variance map into account. The average map informs us the local structure of the human body parts and how it looks like in general. The variance map points out the region where high variation in human poses takes place. Based on both the average and variance information, we create two types of templates, each employs a different design strategy and generates the feature value in different ways. We also utilize the co-existence of strong responses that take place in non-neighboring region pairs to enrich the feature pool. Experimental results show that the proposed method achieves good performance on the INRIA and Caltech-USA benchmark with small feature size, short training time, and less training examples are required.