YOLOv8 multi-scale collaborative and lightweight vehicle-person detection algorithm for intelligent driving

In complex traffic scenarios, vehicle-person target sizes vary greatly, there are many environmental interferences, and the computing power and storage of autonomous driving and roadside edge sensing equipment are limited. To solve such problems, the study proposes a multi-scale collaborative single-stage target detection algorithm, which improves multi-scale feature extraction and fusion capabilities by integrating three modules: spatial-channel synergy, dynamic convolution-warping, and multi-scale dilated attention. for lightweight reconstruction for edge deployment, a lightweight multi-scale collaborative single-stage target detection method is designed. The model introduces a lightweight star backbone network and a mobile inverted bottleneck convolution module to replace the redundant network, and uses an inner wise Intersection over Union Ratio (IoU) loss function to optimize bounding box regression. The average accuracy of the multi-scale collaborative is 87.7% when the IoU threshold is 0.5. The lightweight multi-scale collaborative algorithm compresses the parameter size to 6.83 M. In complex full scenarios, the comprehensive average accuracy reaches 72.7%. On the low-power edge computing platform, the inference speed reaches 22 frames per second. When the batch size is 16, the video memory usage is only 1215 MB. The proposed model breaks the technical barrier between high precision and lightweight in target detection, and provides a reliable solution for the engineering real-time deployment of intelligent traffic sensing systems.