Improving Scan Registration Methods Using Secondary Point Data Channels
MetadataShow full item record
Autonomous vehicle technology has advanced significantly in recent years and these vehicles are poised to make major strides into everyday use. Autonomous vehicles have already entered military and commercial use, performing the dirty, dull, and dangerous tasks that humans do not want to, or cannot perform. With any complex autonomy task for a mobile robot, a method is required to map the environment and to localize within that environment. In unknown environments when the mapping and localization stages are performed simultaneously, this is known as Simultaneous Localization and Mapping (SLAM). One key technology used to solve the SLAM problem involves matching sensor data in the form of point clouds. Scan registration attempts to find the transformation between two point clouds, or scans, which results in the optimal overlap of the scan information. One of the major drawbacks of existing approaches is the over-reliance on geometric features and a well structured environment in order to perform the registration. When insufficient geometric features are present to constrain the optimization, this is known as geometric degeneracy, and can be a common problem in typically environments. The reliability of these methods is of vital importance in order to improve the robustness of autonomous vehicles operating in uncontrolled environments. This thesis presents methods to improve upon existing scan registration methods by incorporating secondary information into the registration process. In this work, three methods are presented: Ground Segmented Iterative Closest Point (GSICP), Color Clus- tered Normal Distribution Transform (CCNDT), and Multi Channel Generalized Iterative Closest Point (MCGICP). Each method provides a unique addition to the scan registration literature and has its own set of benefits, limitations, and uses. GSICP segments the ground plane from a 3D scan then compresses the scan into a 2D plane. The points are then classified as either ground-adjacent, or non-ground-adjacent. Using this classification, a class constrained ICP registration is performed where only points of the same class can be corresponded. This results in the method essentially creating simulated edges for the registration to align. GSICP improves accuracy and robustness in sparse unstructured environments such as forests or rolling hills. When compared to existing methods on the Ford Vision and Lidar Dataset, GSICP shows a tighter variance in error values as well as a significant improvement in overall error. This method is also shown to be highly computationally efficient, running registrations on a low power system twice as fast as GICP, the next most accurate method. However, it does require the input scans to have specific characteristics such as a defined ground plane and spatially separated objects in the environment. This method is ideally suited for outdoor sparse environments and was used with great success by the University of Waterloo’s entry in the NASA Sample Return Robot Challenge. CCNDT provides a more adaptable method that is widely applicable to many common environments. CCNDT uses point cloud data which has been colorized either from an RGBD camera or a joint LIDAR and camera system. The method begins by clustering the points in the scan based on color and then uses the clusters to generate colored Gaussian distributions. These distributions are then used to calculate a color weighted distribution to distribution cost between all pairs of distributions. Exhaustively matching all pairs of distributions creates a smooth, continuous cost function that can be optimized efficiently. Experimental validation of the CCNDT method on the Ford and Freiburg datasets has shown that the method can perform 3D scan registrations more efficiently, three times faster on average then existing methods, and is capable of accurately registering any scans which have sufficient color variation to enable color clustering. MCGICP is a generalized approach that is capable of performing robustly in almost any situation. MCGICP uses secondary point information, such as color, intensity, etc., to augment the GICP method. MCGICP calculates a spacial covariance at each point such that the covariance normal to the local surface is set to a small value, indicating a high confidence matching surfaces, and the covariance tangent to the surface is determined based on the secondary information distribution. Having the covariance represented in both the tangential and normal directions causes non-trivial cost terms to be present in all directions. Additionally, the correspondence of points between scans is modified to use a higher dimensional search space, which incorporates the secondary descriptor channels as well as the covariance information at each point and allows for more robust point correspondences to be determined. The registration process can therefore converge more quickly due to the incorporation of additional information. The MCGICP method is capable of performing highly accurate scan registrations in almost any environmental situation. The method is validated using a diverse set of data including the Ford and Freiburg datasets, as well as a challenging degenerate dataset. MCGICP is shown to improve accuracy and reliability on all three datasets. MCGICP is robust to most common degeneracies as it incorporates multiple channels of information in an integrated approach that is reliable even in the most challenging cases. The results presented in this work demonstrate clear improvements over the existing scan registration methods. This work shows that by incorporating secondary information into the scan registration problem, more robust and accurate solutions can be obtained. Each method presented has its own unique benefits, which are valuable for a specific set of applications and environments.