Occluded Pedestrian Attention Network : An occluded pedestrian Detector

Pedestrian detection performance improves greatly with the development of convolution neural network. However, occlusion remains a challenging issue. To solve the problem, in this paper, we proposed a new pedestrian detector called Occlude Pedestrian Attention Network (OPAN). We use anchor settings, which will highlight the features from the pedestrian region. In addition, spatial-wise attention mechanism and channel-wise are used for regression and classification branch separately and guide the detector to pay more attention to the visible part of the pedestrian. Experiments on Caltech and CityPerson datasets show that the proposed algorithm has better performance than RetinaNet. Compared with RetinaNet, the performance is improved 10.65% and 8.44%. Compared with other latest pedestrian detection algorithms, the performance is also great.