LiDAR Localization by Removing Moveable Objects

In this study, we propose reliable Light Detection and Ranging (LiDAR) mapping and localization via the removal of moveable objects, which can cause noise for autonomous driving vehicles based on the Normal Distributions Transform (NDT). LiDAR measures the distances to objects such as parked and moving cars and objects on the road, calculating the time of flight required for the sensor’s beam to reflect off an object and return to the system. The proposed localization system uses LiDAR to implement mapping and matching for the surroundings of an autonomous vehicle. This localization is applied to an autonomous vehicle, a mid-size Sports Utility Vehicle (SUV) that has a 64-channel Velodyne sensor, detecting moveable objects via modified DeepLabV3 and semantic segmentation. LiDAR and vision sensors are popular perception sensors, but vision sensors have a weakness that does not allow for an object to be detected accurately under special circumstances, such as at night or when there is a backlight in daylight. Even if LiDAR is more expensive than other detecting sensors, LiDAR can more reliably and accurately sense an object with the right depth because a LiDAR sensor estimates an object’s distance using the time of flight required for the LiDAR sensor’s beam to detect the object and return to the system. The cost of a LiDAR product will decrease dramatically in the case of skyrocketing demand for LiDAR in the industrial areas of autonomous vehicles, humanoid robots, service robots, and unmanned drones. As a result, this study develops a precise application of LiDAR localization for a mid-size SUV, which gives the best performance with respect to acquiring an object’s information and contributing to the appropriate, timely control of the mid-size SUV. We suggest mapping and localization using only LiDAR, without support from any other sensors, such as a Global Positioning System (GPS) or an Inertial Measurement Unit (IMU) sensor; using only a LiDAR sensor will be beneficial for cost competitiveness and reliability. With the powerful modified DeepLabV3, which is faster and more accurate, we identify and remove a moveable object through semantic segmentation. The improvement rate of the mapping and matching performance of our proposed NDT, by removing the moveable objects, was approximately 12% in terms of the Root-Mean-Square Error (RMSE) for the first fifth of the test course, where there were fewer parked cars and more moving cars.