Fusion Technology of 3D Point Cloud Map for Objects Classification

Advances in artificial intelligence have led to rapid development in autonomous vehicle technology. Automatic driving vehicles not only use LiDAR to detect distance but also use variety of sensors to detect the surrounding environment that are based on artificial intelligence techniques. This paper presents a neural network method of object classification using color images and LiDAR point clouds in static environment. We use three-dimensional coordinate LiDAR point clouds from KITTI database and convert the point clouds to the polar grid map. The scheme filters out the ground points to reduce the number of point clouds on polar grid map and uses a fast cluster to finish a part of classification. We merge the grids that have points and neighbors into the same cluster by forward and backward sliding window. Then, we use fully convolutional neural network to do classification training using color images in order to get object marked images. Since the output of the full convolutional neural network is an image, the classification can be completed more quickly. Thus, objects can be recognized as pedestrian, vehicles, and plants by combining cluster with deep information and marked image.