全方位 LIDAR に基づく自動運転自動車のための物体識別

抄録

Autonomous vehicles have been developed rapidly in recent years for the purpose of decreasing traffic accidents, reducing traffic jam and increasing human life quality. Environment perception is an important key issue in order to bring such vehicles into reality. Accordingly, a robust object classification is needed in terms of identifying obstacles and providing safe path planning. This paper describes an object classification method based on 3D point cloud data generated by LIDAR. Objects are classified into car, pedestrian, bicyclist and background using AdaBoost and Real AdaBoost classifiers. These classifiers have some advantages such as easy implementation, high performance and low processing time. Features such as the object’s size, shape, intensity and velocity calculated from 3 dimensional coordinates are used as inputs to the classifiers. The proposed strategy is to use decision tree as a weak classifier for AdaBoost and 2-dimension probabilistic distribution for Real AdaBoost. A hug database has been used for training, testing and evaluating the performance of the implemented classifier. The results have verified that the proposed method provides promising accuracy in the range of 85%~95% depending on some object conditions such as appearance, mass and number of available object surfaces.