Research On Decision-making Algorithm Ofintelligent Vehicle Based On Future Spatiotemporal Feature

The decision-making algorithm of intelligent vehicle is based on the location information,perception information and the information of the vehicle itself,which can get the current decision-making quantity under the condition of ensuring the safe and stable driving of the vehicle.In the traditional decision-making method,we need to get the perception and location information of lane line,vehicle and lane line,driving area,and heading angle,and then calculate the decision-making value.This process is complex and difficult to carry out global optimization.The end-to-end decision algorithm is a kind of algorithm which can directly get the decision quantity according to the original information input,such as image.This algorithm can abstract some functions of the sensing module and the positioning module into the algorithm,so as to transform a complex system into an algorithm module.As an end-to-end algorithm,we can optimize the algorithm globally according to the output of the algorithm,so that the system controlled by the algorithm can reach the optimal state,without the need to optimize each pre module separately.Aiming at the shortcomings of intelligent vehicle end-to-end decision-making algorithm,this paper proposes an intelligent vehicle end-to-end decision-making network which integrates future spatiotemporal characteristics and image depth information.The main contents of this paper are as follows:(1)In this paper,an end-to-end intelligent vehicle decision-making algorithm based on future spatiotemporal features is proposed.This algorithm uses conv LSTM based spatiotemporal feature extraction module to extract spatiotemporal feature information from various scales,and integrates the feature maps of various scales to calculate steering wheel angle control value.We use the image sequence from the past to the current time as the input of the network,and use the corner value of the future time to supervise the training of the network,so that the network can predict the corner value of the current and future time.Through experiments,It is proved that Conv-LSTM is superior in space-time feature extraction,and the loss function with future information can improve the precision of the decision model.The RMSE value in Udacity-Challenge-ii is 0.0412,has transcended other existing decision models.(2)Considering the similarity of spatial characteristics and image depth values inthe end-to-end decision-making network of intelligent vehicle,this paper takes the depth estimation method of monocular vision based on depth learning as the pre algorithm of the end-to-end decision-making algorithm,improves the accuracy of the decision-making network by providing regional suggestion information and supplementary depth information to the end-to-end decision-making network,and takes into account that in the real world It is difficult to get dense depth annotation information in the scene.The monocular vision depth estimation network based on migration learning proposed in this paper can alleviate the phenomenon of domain drift by using the method of generating anti training,so that the network can also show better results in the dataset without depth information annotation.(3)In this paper,the end-to-end decision-making algorithm of intelligent vehicle based on future spatiotemporal characteristics is integrated with the depth estimation method of monocular vision based on migration learning.The end-to-end decision-making network of intelligent vehicle based on the depth information and future spatiotemporal characteristics can further improve the decision-making accuracy.When the decision-making model proposed in this paper is deployed to the actual vehicle,this paper proposes a new method In order to solve the output jitter problem of the end-to-end decision network,the output smoothing method of the future turning angle is used.