LPCF-YOLO: A YOLO-Based Lightweight Algorithm for Pedestrian Anomaly Detection with Parallel Cross-Fusion

Abstract

To address the issue of high complexity in current pedestrian anomaly detection network models, which hinders real-world deployment, this paper proposes a lightweight anomaly detection network called LPCF-YOLO (Lightweight Parallel Cross-Fusion YOLO) based on the YOLOv8n model. Firstly, the FPC-F (Fast Parallel Cross-Fusion) module, which incorporates PConv, and the S-EMCP (Space-efficient Merging Convolution Pooling) module are designed in the backbone network to replace C2F and SPPF at various scale branches. Additionally, an ADown module is introduced in the third layer to reduce the computational cost. In the neck network, a Lightweight High-level Screening Feature Pyramid Network (L-HSFPN) is designed to replace the PAFPN structure. Furthermore, the Wise-IoU loss function is employed to enhance the model’s localization performance and generalization ability. The experimental results in the UCSD-Ped1 and UCSD-Ped2 datasets show that, compared to YOLOv8n, the proposed approach reduces parameters by 30.33% and FLOPs by 79.01%, achieving 2.09 M parameters and 1.7 G FLOPs; it also results in a 179.62% increase in FPS to 43.9. Meanwhile, the mean average precision (mAP@0.5) is either maintained (in the UCSD-Ped2 dataset) or slightly improved (in the UCSD-Ped1 dataset).