Research On Object Recognition Based On Lidar And Camera Fusion

With the development of Intelligence transportation,intelligent driving technology has received extensive attention.As the basis of intelligent vehicle driving control,the environment perception technology is the key part of intelligent driving technology researches.The environment perception system of intelligent vehicles is composed of lots of functions.Among them,obstacle recognition is one of the most important component.It is required to accurately classify the obstacles(car,pedestrian and cyclist),obtain precise space information of them and run at real-time.At present,most obstacle-recognition system mainly relies on the environment perception platform composed of Lidar and camera.However,in complex traffic environments,both Lidar and camera are subject to a certain degree of influence due to environmental disturbances,resulting in a decline in the performance of obstacle recognition system,especially the performance on the small-scale obstacles such as pedestrian and cyslist.How to make full use of the advantages of the two kinds of sensors to eliminate environmental interference,improve the performance of obstacle recognition system in obstacle classification and obstacles’ space information acquisition,and maintain high real-time performance is very significant and challenging.In order to solve the above problems,an obstacle recognition technology based on the fusion of Lidar and camera is studied.The main research contents and work of this paper are as follows:(1)The method of obstacle recognition based on mutil-Lidar-feature fusion is studied.This method combine road area,obstacles’ shape and grid distribution as constraint.Meanwhile,multiple geometric and reflectivity features are execrated to train an SVM to recognize obstacles.Experimental results demonstrate that the performance of this method is reliable on motor vehicle recognition while poor on non-motor vehicle recognition.(2)The method of obstacle recognition based on camera is studied.An improved efficient semantic segmentation network(Enet-CRF)is designed which is based on an efficient neural network(Enet).The CRF-RNN network is add in the Enet to improve the network’s sensitive of small-scale information.Experimental results demonstrate that the Enet-CRF can further improve the performance of obstacle-classification especially on small-scale information such as pedestrian and cyclist.(3)An obstacle-recognition framework based on Lidar and camera fusion(Enet-CRF-Lidar)is designed.Lidar and camera are deeply integrated through the progress of the proposed framework,which further improve the performance of the obstacle-classification.The proposed framework can not only accurately classify the obstacles,but also obtain precise space information of them.At the same,the proposed framework can perform at real-time.Experiments show that the proposed obstacle recognition method based on the Lidar and camera fusion has strong reliability and environmental adaptability,which can further improve the performance of the smart vehicles’ environment perception system.