基于多层感知器的端到端车道线检测算法

End-to-end lane detection with Multi Layer Perceptron

针对复杂环境中车道线检测效率低的问题,提出了一种基于多层感知器(MLP)的车道线检测算法(LaneMLP).整个算法主要由全局感知器和局部感知器组成,首先通过逐行分类模型对道路环境图像栅格化,将车道线检测转换为逐行分类任务;分类过程中使用MLP模块作为全局感知器提取车道线的全局语义信息和车道间的结构信息,使用组卷积模块作为局部感知器提取车道线的色彩和位置信息;最后对模型进行结构重参数化设计,以实现训练与推理解耦,在训练与推理过程中使用不同的模块组合,达到提高推理准确率和速度兼顾的目的.在CULane数据集上进行了验证,实验结果表明:在推理速度超过每秒350帧的情况下,准确率达到了76.8%,和SCNN算法相比,准确率提高了5.2%,推理速度也提高了5倍.

Aiming at the low efficiency of lane detection in complex environment,a lane detection algorithm named LaneMLP based on Multi Layer Perceptron(MLP)was presented.The algorithm is mainly composed of global and local perceptron.Firstly,the road environment image is rasterized by the row-wise classification model,the lane detection is converted to the row-wise classification task.In the classification process,the MLP module with residual connections is used as the global perceptron to extract the global semantic information of the lanes and the structure information between lanes.The group convolution module is used as the local perceptron to extract the color and location information of the lanes.Finally,the structure of the model is re-parameterized to decouple the training and inference.Different module combinations are used in training and inference to improve both the accuracy and speed of inference.Experiments on a CULane datasets show that the algorithm is highly competitive in both accuracy and inference speed.When the inference speed exceeds 350 frames per second,the accuracy reaches 76.8%.Compared with SCNN,the processing speed is 5 times faster while the accuracy is 5.2%higher.