Road Environment Detection Based On Multi-source Perception Information Fusion

In recent years,with the rapid development of technologies such as artificial intelligence,internet of things and sensors,Unmanned Ground Vehicle(UGV)with vast prospects of application both in military and civilian fields has become the research hotspot in the academic field at present.Road environment detection is the key and prerequisite of the safe driving of UGV.This paper focuses on the road environment detection based on multi-source perceptual information fusion,and details are as follows:In order to solve the problem of sensor calibration in UGV,this paper realizes the calibration of LiDAR based on the matching of scanning points,and obtains the conversion relation between LiDAR coordinate system and vehicle coordinate system.The calibration of LiDAR and camera is realized by the calibration method based on plane constraint,so the mapping relationship between point clouds and images is obtained,and the spatial alignment of the sensor data is achieved which makes preparations for data fusion.This paper analyzes the installation method of four-layers LiDAR and proposes a new road detection algorithm based on its point cloud distribution.The point cloud distribution feature is used to search for point clouds which are distributed regularly with planar scanning characteristics,and the approximate parallel line segment sets are extracted and fitted.The hypotheses and fitting of road boundary are performed for the envelop point cloud sets of regular point clouds.Smoothing and predicting the road boundary based on time-axis context and detecting obstacles in the road area by point cloud clustering.At finally,the safe drivable road area is obtained by fusing the road boundary and obstacle area.Aiming at the shortcomings of using single LiDAR to detect the road environment,this paper introduces the image to the detection process.A fusion of LiDAR and image road detection algorithm based on the conditional random field(CRF)is proposed.This method transforms the problem of road environment detection into the problem of road segmentation.We construct the CRF model of image and point cloud,and extract the relevant features of image and point cloud to train the random forest classifier.The predictive result of classifier is used for the input of the CRF model,and we obtain the road area in the current driving environment by minimizing the energy function of CRF at finally.Experiments show that this method has high accuracy and reliability.