Research on vision-based multi-task environment perception algorithm for intelligent driving

Existing deep learning algorithms usually process only a single detection task in environment perception, which cannot meet the driving needs of driverless vehicles. To this end, a new multi-task environment perception network is designed, which can simultaneously complete vehicle detection and lane line detection; Taking YOLOv8 as the backbone network and integrating the latest C2f module, the efficient extraction of image features is realized; the neck section adopts BiFPN (Bidirectional Feature Pyramid) structure for better feature fusion and semantic preservation; in loss calculation, αIoU is fused to improve the detection accuracy; based on semantic segmentation, the UCC module is utilized to design an efficient lane line detection branch. Comparison experiments show that the average vehicle detection accuracy reaches 80.0%, and the IoU of lane line detection is 27.10%, which is better than other multi-task perception algorithms in terms of comprehensive performance.