Collaborative object detection and semantic segmentation algorithm integrating lightweight deep learning for complex traffic scenes

Autonomous driving perception requires simultaneous object detection and semantic segmentation under complex traffic conditions while satisfying the memory and latency constraints of edge platforms. To address this requirement, a Lightweight Collaborative Detection and Segmentation Network (LCS-Net) is proposed. The network decouples spatial representation from channel mapping through deeply cascaded lightweight convolutions. It integrates multi-scale context using an improved atrous spatial pyramid pooling module alongside bidirectional gated spatial attention. Furthermore, it balances heterogeneous task convergence through dynamic task weighting and dynamic feature-pixel focal loss. On Cityscapes, LCS-Net achieves 82.53% mAP@0.5 and 76.14% mIoU with 4.19 M parameters, 14.71 G FLOPs, and an inference speed of 91.27 FPS. Compared with YOLOP, it improves mAP@0.5 and mIoU by 6.12 and 5.96 percentage points, respectively, while increasing the inference speed by 34.56%. Compared with YOLOPv2, it reduces the parameter count and FLOPs by 66.02% and 64.36%, respectively, and increases the inference speed by 11.73%. On the BDD100K nighttime subset, LCS-Net maintains 72.19% mAP@0.5, while its peak GPU memory consumption is limited to 1.6 GB. These results indicate that LCS-Net provides an effective accuracy–efficiency trade-off for resource-constrained edge deployment in complex traffic scenes.