Research And Implementation Of Lane Detection Based On Resnet Convolutional Neural Network

Lane detection has always been an important topic in autonomous driving systems.Due to the complexity and changes of the real world,the representation of lane varies from scene to scene.Therefore,lane detection still faces a great challenge.Early lane detection mainly relied on traditional image processing techniques by recognizing its color,shape and other features.However,the accuracy and robustness of this method are not high.In contrast,methods based on deep convolutional neural networks work better.The existing lane detection models are large in size,requiring high hardware computing power and storage space,and are not real-time enough to meet the needs of practical applications.Therefore,the research in this paper focuses on the convolutional neural network-based lane detection method,aiming to improve the real-time performance and applicability of the model.The main research work is as follows:A lane detection method based on row classification and a lightweight attention mechanism is proposed.Using ResNet as the backbone network of the lane detection network can alleviate the gradient disappearance problem due to the deepening of the network layers,using a row classification-based method for lane detection,adding a lightweight attention module to enable the network to better obtain information about the target spatial direction features,and integrating global and local features to improve the model’s ability to detect and recognize lane.On the CULane dataset,the method achieves an F1-measure of 69.9%,an improvement of 1.5 percentage points compared to UFLD.A hybrid anchor mechanism-based lane detection method is proposed to alleviate the magnified localization error problem,using a loss function for sequential classification,which effectively reduces the average localization error,and improves the ResNet-18 backbone network,which effectively improves the efficiency of lane detection,and the network model inference time even reaches 2.1ms,which exceeds the the good performance of 3.1ms in the existing network UFLD.The detection method in this paper was tested on Tusimple,CULane,and homemade datasets.The results show that the method proposed in this paper works well in a variety of scenarios,proving that the algorithm is feasible for lane detection.