Research on Pedestrian Detection Algorithm Based on Improved YOLOv5s

To address the issues of insufficient detection accuracy and an excessive number of model parameters under small targets and occlusion conditions in the pedestrian recognition task, the present study develops a modified YOLOv5s model specifically tailored for detecting pedestrians. Through the replacement of the convolutional layers of YOLOv5s with the lighter Ghost module, the computational volume is reduced and the feature extraction is accelerated; Simultaneously, by adopting the CBAM attention mechanism in the Neck’s feature fusion layer, we significantly improve the performance of network learning ability and improves the detection accuracy; and the SIoU Loss function is adopted in YOLOv5s, which further accelerates the training speed and improve the detection frame localization accuracy. Through this paper comparative experiment, the improved YOLOv5s on the optimized WiderPerson dataset achieves a 5.3% increase in $A P$, achieves a reduction of 2.21 M in the model’s parameter size and reduces the GFLOPs by 46.2%.