Point Cloud Processing Algorithms for Environment Understanding in Intelligent Vehicle Application

Understanding the surrounding environment including both still and moving objects is crucial to the design and optimization of intelligent vehicles. In particular, acquiring the knowledge about the vehicle environment could facilitate reliable detection of moving objects for the purpose of avoiding collisions. In this thesis, we focus on developing point cloud processing algorithms to support intelligent vehicle applications. The contributions of this thesis are three-fold.;First, inspired by the analogy between point cloud and video data, we propose to formulate a problem of reconstructing the vehicle environment (e.g., terrains and buildings) from a sequence of point cloud sets. Built upon existing point cloud registration tool such as iterated closest point (ICP), we have developed an expectation-maximization (EM)-like technique that can automatically mosaic multiple point cloud sets into a larger one characterizing the still environment surrounding the vehicle.;Second, we propose to utilize the color information (from color images captured by the RGB camera) as a supplementary source to the three-dimensional point cloud data. Such joint color and depth representation has the potential of better characterizing the surrounding environment of a vehicle. Based on the novel joint RGBD representation, we propose training a convolution neural network on color images and depth maps generated from the point cloud data.;Finally, we explore a sensor fusion method that combines the results given by a Lidar based detection algorithm and vehicle to everything (V2X) communicated data. Since Lidar and V2X respectively characterize the environmental information from complementary sources, we propose to get a better localization of the surrounding vehicles by a linear sensor fusion method. The effectiveness of the proposed sensor fusion method is verified by comparing detection error profiles.