Occluded Pedestrian Detection Algorithm Based on Attention Mechanism

In light of the situation that it is difficult to accurately detect pedestrians in real scenes due to mutual occlusion, a feature extraction enhanced detection algorithm based on attention mechanism is proposed. Firstly, attention modules are added to learn the relationship between feature channels and the spatial information of feature maps, so as to enhance feature extraction in the visual area of pedestrian targets. Secondly, according to the actual size of pedestrian data, the k-means++ algorithm is used to cluster pedestrian labels, so as to determine the size and proportion of anchors. Distance-intersection over union loss function (DIOULoss) is used to design the loss function of the detector, so that the regression of the detection box pays more attention to the intersection over union between the candidate box and the real box, as well as the center distance between the two boxes. Finally, a new non-maximum suppression algorithm (DSoft-NMS) is presented to preserve more accurate prediction boxes. The proposed method has been tested on CityPersons and WiderPerson datasets, and the results show that the proposed method with a simple network structure has higher detection accuracy in occluded pedestrian detection, which is convenient for subsequent research.