Lane detection algorithm for autonomous vehicle using machine learning

• Main Author: Wu, Guan Jie
• Other Authors: Lyu Chen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/167291
• Institution: Nanyang Technological University

Lane detection has been one of research areas in computer vision for decades. Traditional computer vision algorithms such as Canny edge detector and Hough Transform are commonly used in lane detection model. However, these models only work well under well-conditioned roads with clear lane marks and no occlusion. Hence, there is a need to shift from traditional computer vision algorithms to deep learning method in feature extraction. Convolution neural network (CNN) has been the de facto feature extraction module in computer vision. Spatial CNN (SCNN), the winner of 2017 TuSimple lane detection challenge, propose a special message passing mechanism for long and thin structure. Nonetheless, it can only run at about 20 frames per second (FPS) on NVIDIA GeForce RTX 2080 Ti GPU due to the heavy computational load of the special message passing mechanism. This project explores the possibility of using transformer architecture as the feature extraction module in a lane detection model. This is because the self-attention layers with sufficient heads can express any convolutional layer. The performance of lane detection model based on SegFormer was compared with SCNN in terms of accuracy, F1-measure and FPS. The PyTorch model was converted to TensorRT engine through intermediary ONNX which further boost the FPS of the model during inference. From the experiments, accuracy of SegFormer is higher than SCNN by 0.35% on TuSimple dataset while being 4 times faster than SCNN in inference speed. For CULane dataset, the F1-measure of SegFormer is lower than SCNN by 1.45% but the inference speed of SegFormer is 2 times faster than SCNN in this case. This justifies the trade-off between accuracy and computational load of the model and proven the strong feature representation power of transformer without any incorporation of lane prior.