This paper presents a lightweight pedestrian and vehicle detection algorithm based on YOLO v8 (You Only Look Once v8). Initially, a multi-scale lightweight module called PConv is introduced to reduce both the model’s parameter count and computational demand while preserving its effectiveness, thereby achieving model lightweighting. Subsequently, inspired by the GFPN (Global Feature Pyramid Network) approach, a global feature pyramid is constructed. Through the DySample sampling method, multi-scale fusion extracts feature information, effectively increasing network depth. This enables the neural network to accommodate both shallow and deep semantic information, thereby further enhancing the algorithm’s ability to express feature information. Experimental results on the sliced dataset from bdd100k indicate that, compared to the YOLO v8n algorithm, the mAP0.50:0.95 of the algorithm presented in this paper has increased by 0.4%, while the size of the model has been reduced by 61%. This satisfies the requirements for high detection performance and real-time operation, making it more suitable for deployment and application in real-world production environments.