Reinforced attention method for real-time traffic line detection

The task of traffic line detection is a fundamental yet challenging problem in computer vision. Previous traffic line segmentation models either tend to increase the network depth to enhance the representation ability to achieve high accuracy, or tend to reduce the number of model layers or hyper-parameters to achieve real-time efficiency, but how to trade off high accuracy and low inference time is still challenging. In this paper, we propose a reinforced attention method (RAM) to increase the saliency of traffic lines in feature abstraction, using RAM to optimize the model can achieve better traffic line detection accuracy without increasing inference time. In the RAM processing, we define the line to context contrast weight (LCCW) to represent the traffic line saliency in the feature map, which can be calculated by the ratio of the traffic line energy to the total feature energy. After LCCW calculation, we add a RAM loss item to the total loss in backward processing, and then retrain the model to obtain the new parameter weights. To validate RAM on real-time traffic line detection models, we applied RAM to seven popular real-time models and evaluate them on two popular traffic line detection benchmarks (CULane and TuSimple). Experimental results show that RAM can increase line detection accuracy by 1–2% on the CULane and TuSimple benchmarks, and the ERFNet and CGNet almost reach state-of-the-art performance after the models are optimized by RAM. The results also show that RAM can be applied to the optimization of almost all encoder–decoder-based models, and the optimized models are more robust to occlusion and extreme lighting conditions.