Multi-task traffic scene detection algorithm based on large kernel attention

A network structure based on a combination of attention mechanism and large convolution kernel is proposed to address the problem of low detection accuracy of traditional multi-task traffic scene detection algorithms, which is used to perform vehicle object detection, drivable area detection and lane line detection tasks. Firstly, this paper proposes a backbone network that fuses large kernel attention mechanism and ELAN structure to improve the overall recognition accuracy of the model. Secondly, by fusing large convolution kernel attention mechanism and multi-scale information interaction mechanism, this paper designs a segmentation enhancement module to improve the model perform lane line segmentation task. Finally, by combining the characteristics and advantages of Focal loss and Tversky loss, this paper combines them as the loss function of segmentation tasks to solve the problem of sample imbalance in segmentation. The experimental results show that on BDD100K dataset, compared with traditional convolutional neural network algorithms, the recall rate of object detection of the proposed network reaches 91.1%, the accuracy of lane line detection reaches 84.0%, and the mean intersection over union of drivable area detection reaches 92.7%, which are significantly improved. Meanwhile, the proposed algorithm achieves a detection speed of 154fps on RTX3090, meeting the real-time requirements.