Research On Obstacle Detection In Railway Station Yard Based On Fusion Of Lidar And Camera

With the increase of train operating mileage in China,the number of railway station yards has been increasing and dispatching tasks have been becoming heavier.The railway station yard is characterized by large amount of trains,complex track lines and random movement of staff,which poses great theratens to train dispatching.Therefore,it is of great significance to quickly and accurately detect the obstacles in front of the train in railway station yards.Most of the existing object detection in railway scenes is based on images.however,the process of acquiring images through cameras depends on environmental factors such as weather and lighting,and image-based detection results lack orientation information,which cannot meet the requirements of railway station scene perception.Aiming at addressing above problems,this paper proposes a object detection method based on the fusion of lidar and camera,for identifying obstacles in front of the train in railway station yards.The proposed method can the advantages of both lidar and camera to obtain the category and orientation information of objects,and then judge whether the objects are obstacles according to the orientation information,which provides more valuable information for the safety scheduling of trains.The main research contents of this paper are as follows:Considered the particularity of railway station yard,this paper proposes a 3D point cloud target detection approach based on knowledge rule matching.In order to improve the timeliness of the proposed approach,the original 3D point cloud is filtered before the point cloud segmentation,so as to achieve the purpose of point cloud downsampling.Afterwards,track point cloud is fitted using a manner of straight line,and then the background point cloud is segemented using the fitted straight line.Secondly,the clustering algorithm is used to cluster the segmented3 D point cloud,and the features of the independent point cloud clusters obtained after clustering are extracted.Finally,the object detection of 3D point cloud is performed matching the extracted features with the designed knowledge rules.Experimental results show that the proposed approach has high accuracy.Aiming at timeliness existed in object detection algorithm,this paper uses the pruning principle to lighten the model based on the traind YOLOv4,which improves the timeliness of the algorithm.Based on the image data collected on site,the YOLOv4 algorithm is used to train the image-based target detection model.Then the γ coefficients in the batch normalization layer of the model are sparsely trained,channels and layers can be pruned according to the distribution of the γ coefficients in the model.Finally,the pruned model is retrained to improve the accuracy of the model.The experimental results show that under the condition of ensuring the accuracy of target detection,the speed of the model is greatly improved,which can better meet the realtime requirements.On the basis of above results,a decision-level fusion method is designed to detect obstacles in railway station yard based on the fusion of lidar and camera.Given that different sensors have different spatial poses and data collection frequencies,the camera’s internal parameter calibration is firstly implemented using collected checkerboard data.Then,the extrinsic parameter matrix between the lidar and camera is calcualted by the feature consistency principle of the object in 3D point cloud and image.Moreover,the time alignment between the lidar and camera is achieved through the working frequency relationship between them.Furthermore,the feature points in 3D point cloud are projected to image,and a fusion algorithm is designed to enable the fusion of detection results obtained by both 3D point cloud and image,which can judge whether the object is an obstacle according to the fusion results.Finally,the effectiveness of the propsoed fusion method is verified by case studies.The experimental results show that the propsoed fusion method can achieve better detection performance than single sensor-based mtehod,and further improve the accuracy of scene perception for railway station yard,which has high practical value.