A 3D Obstacle Classification Method in Point Clouds Using K-NN

Object classification and recognition is a crucial function for unmanned ground vehicle (UGV) to realize smart environment perception and obstacle avoidance. Based on the advantage features of fast collection, high-precision, and wide-covered, Light Detection and Ranging (LiDAR) is widely equipped on UGV to collect environment information. To analyze the sensed LiDAR point cloud, this paper developed an obstacle classification system using the k-Nearest Neightbour (k-NN) algorithm to identify objectsto its corresponding categories. Before object classification, point cloud existed in the outdoor environment is segmented into several sub-point-clouds according to their space distribution. This way, the whole entire scene is divided into separated obstacles, which are the pre-process of proposed object classification method. Using the segmented object point cloud, geometry features are extracted out as the judgment basis to recognize obstacle types. Combined with the object features, the manually marked object category are stored together as the training datasets of the k-NN model. When a new testing data is fed into the k-NN mode, the object type is obtained through counting the most object types in nearest k training datasets.