Fusion of modern and tradition: A multi-stage-based deep network approach for head detection

Abstract

Detecting humans in video is becoming essential for monitoring crowd behavior. Head detection is proven as a promising way to realize detecting and tracking crowd. In this paper, a novel learning strategy, called Deep Motion Information Network (abbr. as DMIN) is proposed for head detection. The concept of DMIN is to borrow the traditional well-developed head detection approaches which are composed of multiple stages, and then replace each stages in the pipeline into a cascade of sub-deep-networks to simulate the function of each stage. This learning strategy can lead to many benefits such as preventing many trial and error in designing deep networks, achieving global optimization for each stage, and reducing the amount of training dataset needed. The proposed approach is validated using the PETS2009 dataset. The results show the proposed approach can achieve impressive speedup of the process in addition to significant improvement in recall rates. A very high F-score of 85% is achieved using the proposed network that is by far higher than other methods proposed in literature.