MLFFCSP: a new anti-occlusion pedestrian detection network with multi-level feature fusion for small targets

Pedestrian detection relying on deep convolution neural networks has achieved significant progress. However, the performance of current pedestrian detection algorithms remains unsatisfactory when it comes to small targets or heavily occluded pedestrians. In this paper, a new anti-occlusion video pedestrian detection network with multi-level feature fusion named MLFFCSP is proposed for small targets and heavily occluded pedestrians. In the proposed network, the pyramid convolutional neural network PyConvResNet101 is used as backbone to extract features. Then, the shallow and deep features are fused at multiple levels to fully obtain the shallow location information and deep semantic information. In order to improve the robustness of the model, data augmentation is also implemented via random erasing on the training data. Experiments are carried out on Caltech and Citypersons datasets, and the log-average miss rate is used to evaluate the performance of the model. The results show that the performance of MLFFCSP is better than other pedestrian detection algorithms in the case of small targets and serious occlusion.