Deep Learning Based Collision Avoidance Strategy For Unmanned Ground Vehicles

Since the first unmanned ground vehicle(UGV)turned up during the 1960s,UGV technology has made great breakthroughs,emerged a lot of advanced technologies in this area.This paper focused on the current state in UGV research,summarizing its important achievements and progress.At the same time,this paper investigated the state-of-the-art developments in deep learning algorithms and vehicle collision avoidance strategies,and trying to apply them together to improve the UGV collision avoidance strategy.Combined the strengths of Lidar in data collecting with the characteristics of deep learning algorithm in unsupervised feature extraction.By analyzing and processing on the point cloud data collected from the on-vehicle Lidar,extracting the information of obstacles around the car.Furthermore,using deep learning networks to classify the obstacle scenarios,for the purpose of finding the best solutions to copy that situation,providing guidance for the UGV in collision avoidance.The creative achievements in this paper including as follows,(1)For the large number of Lidar datum collecting from experiments,proposed a totally new algorithm——obstacle information retrieval algorithm to distinguish the obstacles.Firstly,mapping each frame of the data collected to a 2-D RGB figure,generating multi-frames images with timing relationships.Secondly,analyzing the relative movement of the spatial point by comparing the changes in G values of corresponding z values of two sequential frames,getting the contour of obstacle information around the car.(2)Focusing on the 360° environmental conditions around the car based on the data collected.Treating the area that the lidar can reach as a whole.Combining the big data technologies with deep learning algorithms to help the data analysis and decision making during the UGV collision avoidance strategy.Using the advantages of computer in massive data processing,looking for a more comprehensive and more efficient collision avoidance strategy than manual designs.(3)Designed a deep learning neural networks,using the corresponding relationships between the 'scanning field image'and the 'driver behaviors'to describe human drivers' driving habits.Sampling huge number of human driving behaviors data,put the data pre-processed before into the deep neural networks as the input data,then train the data and tuning parameters by layer-wise method.Extracting the information about the changes in acceleration and direction of the vehicle as the output of the network from large amount of driving data,which can be used as an important reference for driving strategy decisions.For the issue of blind spots of 64E lidar,in further study,data fusion can be applied with other on-vehicle lidars and the cameras,cooperating with object recognition,environmental perception to optimize the UGV driving strategy.