Occluded Pedestrian Detection and Image Recognition with Multi-Attention Context Networks

For the complex traffic road scenarios where the occluded pedestrians are difficult to be detected by detectors, Multi-Attention Context Network (MACNet) is proposed, aiming to use contextual information and attention mechanism to handle the occluded pedestrians. Firstly, we add the multi-attention context module to make the detector obtain richer contextual information and use its attention mechanism to learn different occlusion patterns. On this basis, add trainable parameters to combine the global context module with the multi-attention context module to establish an adaptive mutual supervision mechanism to further improve the feature extraction of obscured pedestrians. Finally, unreasonable samples and too small positive samples are ignored in the network training process to reduce the negative impact of such samples on the network training. Experimental results show that the proposed method reduces the detection miss rate in different scenarios, and the improvement of pedestrian detection miss rate in heavy occlusion is more obvious.