Sensor fusion is our future

Anomaly detection of vehicles requires well-defined surrounding vehicles’ positions for monitoring the movement and predicting other vehicles behaviors. To propose the exact location of other vehicles, this paper explores the ways to accurately estimate the lateral distance between mid-low point on the rear of the vehicle and adjacent lane by detecting lane markers and objects with orientation information. Both are strongly dependent on advantages of Camera and LiDAR sensors and rely on Deep Neural Network. To estimate vertical distance of other vehicles about lanes, we detect the lanes and objects with image-based deep learning model and point cloud-based deep learning model respectively. Additionally, we focus on the projection process which puts 3D point clouds on the image domain to merge the detection results from different dimensions. By putting matching process in the projection step, it is able to find 3D lanes corresponding to 2D lane detection results, so getting the lateral distance between target points on vehicles and lanes is possible with differential relation of 3D object detection result and lane detection result on 3D. In order to present the estimation performance of our method, this paper considers error between the rear center point of the 3D objection box and the same target point on the vehicle identifiable on the actual point cloud map, and calculates the error that occurs in the process of matching the lane data detected in 2D to 3D.