| The three components of intelligent driving technology are perception,decisionmaking and control.Only the reasonable division and cooperation of the three systems can realize safe and reliable intelligent driving.Perception system is the most basic part of intelligent driving system.It mainly obtains the surrounding environment information in real time and dynamically,so as to provide necessary guarantee for decision-making and control.As one of the key technologies in the environmental perception module,the significance of freespace segmentation for automatic driving lies in that on the one hand,it can provide the relative position of the current vehicle on the road to ensure that the vehicle runs in a safe area;On the other hand,it provides effective and reliable perceptual information for the decision-making and control module to form a more perfect safe driving system.In this thesis,lane detection and freespace segmentation in intelligent driving system are taken as the research content,and are realized its single task respectively.In view of the situation that the lane line is easy to be blocked and the illumination changes in urban roads,the lane detection method based on row anchor is established as the basic framework.The model is training on the open source dataset BDD100 K,and using the tuning skills to improve its performance.Add the self built dataset to finetune the model and improve the generalization.Verify the effectiveness of the lane detection model according to the indicators on the self built test set and the visual results.According to the actual scene requirements,the classification standard of freespace based on road material is established.Based on the network architecture of bilateral segmentation network,the network structure is improved by changing the detail branch and semantic branch into Rep VGG structure respectively.The model is trained and optimized based on open source datasets Cityscapes and BDD100 K and self built dataset to realize pixel level freespace segmentation.Through the analysis of the defects of the existing single task technology of lane detection and freespace segmentation,a multi task learning network MTNet with unified sharing of bottom parameters taking the hard parameter sharing model as the theory is proposed in this thesis,which is divided into different branches in the deep layer of the network to realize two tasks respectively,and realizes the network architecture based on pytorch deep learning framework.By building the dataset and completing the data preprocessing work such as label conversion and data expansion,it is input into the multi task network,and by adjusting the hyperparametric optimization model,the performance of the final model is evaluated through the indicators and the visualization effect in different weather and different periods.The experimental results show that the indicators of lane line detection branch and driveable area segmentation branch of multi task model MTNet are improved to a certain extent compared with the single task model,and can meet the expected requirements in real-time and robustness.To sum up,t a scheme of lane detection and freespace segmentation based on multi task learning is proposed and experimental research is carried out in this thesis,which has certain practical significance for the realization of automobile intelligence. |
