Lane Detection Based On Semantic Segmentation

Lane detection provides important path information for vehicles,and thus is an important part of intelligent driving environment perception.In recent years,the application of deep learning has significantly improved the performance of lane detection.However,it is still challenging to detect lane robustly and efficiently in complex environments.In this thesis,a lane detection algorithm based on lightweight semantic segmentation is proposed and transplanted on the small,low-power embedded SE3 AI platform.The main research contents of this thesis are as follows:(1)Lane detection algorithm based on SFNet semantic segmentation network.Generally,in the process of feature fusion,semantic segmentation networks cause the loss of object edge information due to the semantic misalignment of feature maps of different resolutions.The Feature Alignment Module(FAM)based on semantic flow in SFNet can effectively mitigate this issue.In this thesis,SFNet is used to refine the lane segmentation boundary,which effectively improves the segmentation accuracy.To solve the imbalance between positive and negative samples and the imbalance between classes in the dataset,the online hard example mining algorithm(OHEM)is used to improve the original weighted cross-entropy loss function,which strengthens the network’s learning ability for few samples.To further speed up,lightweight backbone DF2 is used to extract features.Compared with previous research work based on ICNet,on the Baidu Apollo Scape dataset containing multiple road scenes,the lane detection algorithm proposed in this paper increases the accuracy from 67.1% m Io U to 75.2% m Io U,and the speed from 30 FPS to 76 FPS.(2)Post-processing based on improved RANSAC.Due to the influence of the image perspective effect,the shape of the lane is irregular,resulting in a large number of invalid iterations in the RANSAC fitting algorithm,and the algorithm is inefficient.Different from the traditional algorithm,the lane candidate regions are uniformly divided to make the algorithm iterate efficiently.According to the characteristics of lane fitting,the iteration termination condition and fitting method are improved,which significantly improves the fitting speed.Compared with the traditional RANSAC,the improved RANSAC algorithm reduces the time cost per frame from 2.56 s to 0.086 s.In the end,the accuracy of the lane detection algorithm on the Baidu Apollo Scape dataset can reach91.7%.(3)Transplantation of lane detection algorithm based on the embedded SE3 AI platform with small size and low power consumption.In order to solve the problem that the SE3 AI platform does not support some complex deep learning operators during the transplantation process,the supported operators are used to replace their functions.In addition,the appropriate image input batch size can give full play to the computing resources of the hardware.Therefore,the model calculation graph is optimized through experiments to improve the inference speed of the model on the SE3 AI platform.Finally,the inference speed of the lane detection algorithm on the SE3 AI platform is increased from 12 FPS to 19 FPS.