Robust outlier rejection methods for vision-based SLAM in dynamic environments

Abstract

This thesis explores various methods to solve the problem of simultaneous localization and mapping(SLAM) in dynamic environments using vision sensors. Most of SLAM algorithms assume that the observed features or landmarks are in a stationary state without movement. However, in the real environment, there are various dynamic objects, such as dynamic objects like people and cars, and temporary static objects which are static during observation but move when they are out of sight, triggering false positive loop closings. In particular, vision sensors typically have a narrow field of view and have inaccurate distance information. Therefore, it is more affected by dynamic objects. In order to deal with dynamic and temporarily static objects, a method that can solve those problems from the perspective of local and global optimization of SLAM is required. Therefore, this thesis is mainly composed of the method for the local optimization method to deal with dynamic objects and the method for the global optimization to deal with static objects. The first part of this thesis deals with the method of dealing with dynamic objects in local optimization. In general, sensor systems usually used in real environments include not only vision sensors but also wheel odometry or inertial sensors. These auxiliary sensors have the characteristic of being stable for short travel distances and are not affected by the change of the feature point by a moving object. In order to reflect the information of those sensors, the sensor data obtained in the local window are grouped first. The first method proposed in this part is a local optimization method using an RGB-D sensor and wheel odometry. Sensor data(nodes) observing the same scene are grouped through feature point tracking. Then, feature matchings are performed within the group. Among these matches, matches derived from a dynamic object are pruned by the Mahalanobis distance with respect to the wheel odometry, and only the matches from a static object are retained and used for local optimization. Through this method, robust and accurate local localization and mapping can be done. Second, a local optimization method using a vision sensor and an inertial sensor is proposed. In the sliding window of the visual-inertial algorithm, weights are given to each feature point. Then, the regularization factor that determines whether the feature point is dynamic and the momentum factor for consistency by reflecting tracking information are added to eliminate the influence of the dynamic features. Finally, the method using Graduated Non-Convexity instead of the two factors is proposed to achieve the real-time performance of the robust visual inertial algorithm. Since local optimization was performed in the previous part, the second part of this thesis deals with the methods that can deal with temporarily static objects during global optimization using locally optimized trajectory and mapping information. As a first method, feature matching between groups, not nodes, is performed to shorten the time by reducing the number of trials of matches. Constraints from the matched results are grouped according to the displacement of the group's reference node, and the groups of these constraints called hypotheses are pruned using the Mahalanobis distance or removed by using a regularization factor during optimization. Finally, this thesis is concluded by analyzing the limitations of the proposed methods and suggesting future research directions to solve them.