Research on Pedestrian Detection Based on Improved YOLOv7-Tiny Algorithm

This paper presents a multi-level receptive field fusion YOLOv7-tiny algorithm to mitigate the high false detection and miss rates in pedestrian detection tasks, which are primarily due to the limited effective information of small target pedestrians and pedestrian occlusion. Firstly, to address the issue of a single receptive field in feature extraction and fusion in the original YOLOv7-tiny algorithm, we propose the RFB-ELAN method. This method employs multi-branch convolution to fuse receptive fields of varying sizes, thereby effectively identifying pedestrians across different scales. Subsequently, addressing the problem of the loss function becoming ineffective due to the high sensitivity of small-scale pedestrians to positional deviations in traditional metrics based on Intersection over Union (IoU). We convert the IoU of prediction boxes and ground truth boxes into the similarity of two Gaussian distributions, and proposes a loss function (NGWDI Loss) based on the normalized Gaussian distance and CIOU fusion. We conducted algorithm comparison and ablation experiments on the CityPersons dataset. The experimental results demonstrate that, compared to the baseline method, our proposed method improved the Average Precision (AP) by 6.02%, the Miss Rate (MR−2) by 4.62%, and the Recall by 13.56%. These results indicate superior performance of the proposed method.