Small-Scale Pedestrian Detection Using Fusion Network and Probabilistic Loss

Small-scale pedestrian detection is a challenge. The main issues are as follows: 1) Troubled by their small scale, it is difficult to extract features effectively; 2) During the detection process, it is easily disturbed by background noise such as inter-class occlusion and intra-class occlusion, leading to missed or false detection; 3) The current widely used IoU measurement method is very sensitive to the position deviation of small objects, which seriously reduces the detection performance. To address these problems, we improve YOLOv5 structure by integrating Non-Local and Convolution structures, building a new feature extraction module called ResNet-Conv&NonL, combined with the ResNet structure. This module was then integrated into the backbone of YOLOv5 for better image feature extraction. In addition, we developed a novel model to measure the similarity between bounding boxes, which are embedded in the loss function of the YOLOv5 structure to replace the normal IoU measurement. Experiments on a self-made dataset and a combined dataset from Caltech and CityPersons show the feasibility of the proposed network structure. Results demonstrate the feasibility of the improved network structure is superior to the original method because it increases average precision by 6.0% compared to the original one.