Obstacle Detection And Moving Object Tracking Based On Spatial-Temporal Fusion Of LiDAR

LiDAR has a wide scan range and high accuracy and is widely used to detect obstacles and track objects in the field of auto-driving.Traditional LiDAR sensing methods usually use the single frame point cloud to sense the environment.Because of the sparsity of LiDAR,single-frame sensing method has a poor performance in detecting distant obstacles.By using spatial-temporal fusion,we can achieve a great improvement in LiDAR sensing.Our spatial-temporal fusion method is developed in the data layer and the target layer respectively.In the data layer,we use the dense point cloud to detect obstacles with the help of pose provided by registration.In the target layer,we track the dynamic objects and eliminate the error caused by them in multi-frame detection.our contributions is as follows:1.Realize real-time registration of multi-frame point cloud in data layer.We extract feature points from the original point cloud according to curvature.Cost function is constructed according to point-line distance and point-surface distance between feature points.After two-stage matching of inter-frame registration and map registration,we can adjust the pose from rough to fine.Compared with traditional ways,our method involves a small number of LiDAR points,and it has a good performance both in accuracy and speed.2.Detect obstacles using registered dense point cloud.Using the pose obtained by registration,we can transform historical point cloud to current frame.After projecting multi-frame point cloud to grid map,we can classify every grid to different attribute by the height distribution.Compared with single-frame detection,multi-frame detection can achieve a further sensing distance after spatial-temporal fusion.3.Kalman filter is used to track moving objects in the target layer.Dynamic objects can cause error in multi-frame detection.Firstly,cluster obstacle grid and extract contour to get candidate targets.Then,match targets of different frame using nearest neighbor search with the help of pose provided by registration,and screen out moving targets.Finally,Kalman filter is used to track moving objects and and estimate their speed.Compared with traditional ways,our method does not need extra motion sensor and has a high cost performance ratio.