A Channel-Cascading Pedestrian Detection Network for Small-Size Pedestrians

At present, there are several new challenges for multi-scale pedestrian detection in wide-angle field of view, especially small-size pedestrians. So the problem is how we can detect pedestrians efficiently and accurately with limited resources in wide-angle field of vision. In this work, we propose a Channel-Cascading pedestrian detection network for small-size pedestrians. In combination with the two-stage idea of Faster-RCNN in our detector, the optimized network was applied and the regional proposal network was improved. We propose a novel feature extraction network as optimized network, which we call the “Channel-Cascading Network” (CCN), that fuses information between channels by progressive cascading strategy and adapts our idea to other network designs. The experimental results show that our detector performs better for small-size pedestrians, it not only the precision of pedestrian detection in wide field of view is greatly improved especially small-size pedestrian, but also the speed is accelerated.