Occluded Pedestrian Classification Using Gradient Patch and Convolutional Neural Networks

Occlusion handling has been an important topic in pedestrian recognition. This paper proposed new approach for occlusion handling by Gradient Patch and Convolutional Neural Network (CNN). There are several researches of occlusion handling use parts annotations or manual labeling of body parts. However our method is learning partial features without any prior knowledge. Our model is trained parts detector with multiple of partial features that selected by gradient patch. Gradient patch compute the orientation of the edge in sub-region and find the extra partial features along the edge directions. Our experiments represented the effectiveness of Gradient Patch for occlusion handling in the INRIA and Daimler pedestrian dataset.