Pedestrian detection using RetinaNet with multi-branch structure and double pooling attention mechanism

Pedestrian detection technology, combined with techniques such as pedestrian tracking and behavior analysis, can be widely applied in fields closely related to people's lives such as traffic, security, and machine interaction. However, the multi-scale changes of pedestrians have always been a challenge for pedestrian detection. Aiming at the shortcomings of the traditional RetinaNet algorithm in multi-scale pedestrian detection, such as false detection, missed detection, and low detection accuracy, an improved RetinaNet algorithm is proposed to enhance the detection ability of the network model. This paper mainly makes innovations in the following two aspects. Firstly, in order to obtain more semantic information, we use a multi-branch structure to expand the network and extract the characteristics of different receptive fields at different depths. Secondly, in order to make the model pay more attention to the important information of pedestrian features, double pooling attention mechanism module is embedded in the prediction head of the model to enhance the correlation of feature information between channels, suppress unimportant information, and improve the detection accuracy of the model. Experiments were conducted on different datasets such as the COCO dataset, and the results showed that compared with the traditional RetinaNet model, the model proposed in this paper has improved in various evaluation indicators and has good performance, which can meet the needs of pedestrian detection.