摘要
行为规划是无人车驾驶的关键技术。基于视觉的行为规划提供了廉价的解决方案,但是由于道路图像分布的复杂性和图像计算量的庞大,使得无人车行为规划系统的设计变得困难。针对这一问题,提出一种基于神经网络的无人车行为规划系统,该系统首先将道路图像降维编码,并将编码空间约束为正态分布空间,再在编码空间中实现对无人车的行为控制。实验结果表明,该系统解决了直接在图像空间建模和计算复杂的困难,并能够较好地对无人车进行行为规划和有效地预防错误控制的发生。
Motion planning is the key technology of unmanned vehicle driving.Vision based motion planning provides an inexpensive solution.However,due to the complexity of the road image distribution and the huge amount of image computation,it is difficult to design a motion planning system of the unmanned vehicle.To solve this problem,a motion planning of unmanned vehicle system based on heterogeneous deep learning is adopted.Firstly,the dimension of the road image is reduced and the images are encoded.The encoded space is restricted to normal distribution space.Then unmanned vehicle motion control is realized in the encoded space.The experimental results show that the system solves the difficulty of modeling and computing complex directly in image space and can plan the motion of unmanned vehicles preferably while preventing the occurrence of error control.
作者
兰潇根
石朝侠
LAN Xiaogen;SHI Chaoxia(Nanjing University of Science and Technology, Nanjing 210094)
出处
《计算机与数字工程》
2019年第7期1635-1639,1737,共6页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:61371040)资助
关键词
无人车
行为规划
人工神经网络
变分自编码
unmanned vehicle
motion planning
artificial neural network
variational autoencoder
