Design of a Vehicle Trajectory Prediction System

Le résumé fourni par la source

With the increasing advancement of science and technology, the popularity of automobiles in my country has gradually increased. It's not difficult to realize that our daily life and productions tend to be autonomous and not only that autonomous vehicles are not strange to us. Although completely autonomous driving which is the top level of self-driving is still far from us. Partially self-driving or driving assistance is much nearer to our life. Self-driving car has been facing many problems, like the study of machine learning and artificial intelligence not being perfect firstly. And the algorithms still need to be improved. Besides, the complexity of traffic environment adds the difficulties to it. Secondly, when it comes to deployment, the computing power on a single vehicle is not enough to deal with the predicting tasks and the laws and policies are not mature. Especially at the moral level, it's hard for people to fully accept self-driving. For the prediction of vehicle trajectories, it's the key to recognize the driving intention and lane changing prediction means a lot to it. And this paper is mainly on the prediction of lane changing. Deep learning is deeper and has stronger ability to express than usual machine learning, which is getting more popular in most areas. And the recurrent neural network is good at dealing with time sequences problem and has a great contribution to machine translation. While the gradient problems put limits on the RNN especially when dealing with longer sequences. Therefore, the LSTM is proposed to solve gradient obstacles and has great performance on predicting tasks. Based on the existing basics, the LSTM buy in the door control unit to better learn historical information and productivity, performance gets optimized with the attention mechanism. We proposes a LSTM encoder decoder based model and uses social pooling to gain more information about surrounding vehicles. After knowing about some data sets, we finally choose the open data set NGSM and its US-101 and I–80 data.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.