Pedestrian detection via deep segmentation and context network

Abstract

For pedestrian detection, many deep learning approaches have shown effectiveness, but they are not accurate enough for the positioning of obstructed pedestrians. A novel segmentation and context network (SCN) structure is proposed that combines the segmentation and context information for improving the accuracy of bounding box regression for pedestrian detection. The SCN model contains the segmentation sub-model and the context sub-model. For separating the pedestrian instance from the background and solving the pedestrian occlusion problem, this paper uses the segmentation sub-model for extracting pedestrian segmentation information to generate more accurate pedestrian regions. Considering that different pedestrian instances need different context information, this paper uses context regions with different scales to extract context information. For improving the detection performance, this paper uses the hole algorithm in the context sub-model to expand the receptive field of the output feature maps and connect the multi-channel features with the skip layer. Finally, the loss functions of the two sub-models outputs are fused. The experimental results on different datasets validate the effectiveness of our SCN model, and the deeply supervised algorithm has a good trade-off between accuracy and complexity.