Abstract:Monitoring vehicular road traffic is a key component of any autonomous driving platform. Detecting moving objects, and tracking them, is crucial to navigating around objects and predicting their locations and trajectories. Laser sensors provide an excellent observation of the area around vehicles, but the point cloud of objects may be noisy, occluded, and prone to different errors. Consequently, object tracking is an open problem, especially for low-quality point clouds. This paper describes a pipeline to integrate various sensor data and prior information, such as a Geospatial Information System (GIS) map, to segment and track moving objects in a scene. We show that even a low-quality GIS map, such as OpenStreetMap (OSM), can improve the tracking accuracy, as well as decrease processing time. A bank of Kalman filters is used to track moving objects in a scene. In addition, we apply non-holonomic constraint to provide a better orientation estimation of moving objects. The results show that moving objects can be correctly detected, and accurately tracked, over time, based on modest quality Light Detection And Ranging (LiDAR) data, a coarse GIS map, and a fairly accurate Global Positioning System (GPS) and Inertial Measurement Unit (IMU) navigation solution.

This paper proposes a method for tight integration of IMU (Inertial Measurement Unit), stereo <small>VO</small> (Visual Odometry) and digital map for land vehicle navigation, which effectively limits the quick drift of <small>DR</small> (Dead Reckoning) navigation system. In this method, the <small>INS</small> provides the dynamic information of the land vehicle, which is used to predict the position and attitude of cameras in order to obtain the predicted pixel coordinates of features on the image. The difference between the measured and predicted pixel coordinates is used to reduce the accumulated errors of <small>INS</small>. To implement the proposed method, an Extended Kalman filter (<small>EKF</small>) is first used to integrate the inertial and visual sensor data. The integrated solution of position, velocity and azimuth is then applied by fuzzy logic map matching (<small>MM</small>) to project the vehicle location on the correct road link. The projected position on the road link and the road link azimuth can finally be used to reduce the dead reckoning drifts. In this way, the accumulated system errors can be significantly reduced. The testing results indicate that the horizontal <small>RMSE</small> (root-mean-square-error) of the proposed method is less than 20 meters over a traveled distance of five kilometers and the relative horizontal error is below 0.4 percent.