Lightweight Pedestrian Detection Algorithm Based on Improved YoLov8s

To address the high computational complexity and parameter redundancy in pedestrian detection models, this paper proposes a lightweight YOLOv8s-based algorithm optimized for resource-constrained devices. The improvements include three key components: (1) Integration of RepViT blocks into the backbone network to replace traditional bottlenecks, where the Squeeze- and-Excitation (SE) attention mechanism is selectively removed from shallow layers to reduce redundancy while retaining it in deeper layers for enhanced feature representation. (2) Replacement of standard convolutions with Group Spatial Convolution (GSConv) to minimize parameters and computational costs. (3) Introduction of the Shape-IoU loss function, which incorporates shape and scale alignment to improve bounding box accuracy for irregular pedestrian contours. Experiments on the WiderPerson dataset demonstrate that the proposed model achieves S9.5<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> [email protected] and 63.7% [email protected]:0.95, comparable to the original YOLOv8s, while reducing parameters and computational effort by 36.6% (to 7.1 MB) and 32.6% (to 19.4 GFLOPs), respectively. The model also outperforms YOLOv5n, YOLOv6n, and YOLOv8n in recall (8O.8%) and precision (S7.2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>), achieving an optimal balance between efficiency and accuracy. These advancements enable practical deployment in intelligent transportation, surveillance, and edge-computing scenarios.