Search

This paper presents a non-iterative pose-graph optimization algorithm for fast 2D simultaneous localization and mapping (SLAM). The graph-SLAM approach addresses the SLAM problem using a factor graph structure. For a pose-only SLAM problem, landmarks are not explicitly modeled and are not a part of the SLAM problem. Conventional pose-graph optimization methods minimize the error by an iterative local linearization process. The proposed method reformulates the pose-graph optimization problem as a combination of two linear least-squares problems. The position and angle term in a pose vector are optimized separately, and the iterative linearization process is removed. Due to an approximation in the reformulation of the pose-graph optimization problem, there is a tradeoff between the accuracy and the computational complexity. The simulation is conducted to demonstrate the efficiency of the proposed method. For comparison, the conventional manifold based pose-graph optimization algorithm is implemented. The results of simulations which optimized 1,079 poses show that the proposed method is more than 25 times faster than the conventional method. However, the localization accuracy is approximately 10% lower than the conventional method.

This paper presents the performances of a corner detector and a region detector in vision-based Simultaneous Localization and Mapping (SLAM). In vision-based SLAM, a feature detection process for data-association is the first step. Therefore, the repeatability of a feature detector is important. Feature detectors are generally categorized as the region detector and the corner detector. There are conflicting research results for the repeatability of corner detectors and region detectors. This paper shows SLAM results using a corner detector and a region detector. Extended Kalman Filters (EKF) are used for implementing the two SLAM methods. The localization errors are evaluated quantitatively. From the experimental results, it is demonstrated that the corner detector is more efficient than the region detector for the work performed in this paper.

This paper presents a new localization method by the data fusion of vision sensor and motion sensor measurements, which can be used for mobile robots. The developed system estimates the orientation, position and velocity of a mobile robot. This is achieved by using data from a camera, robot wheel encoders, an accelerometer, and a gyroscope. The developed system is a multi-rate sampled data system. For the accurate estimation of robot position and velocity, the developed method detects the slip of robot wheels, by comparing the data from the encoders and the accelerometer. The developed method estimates the robot orientation and position by using an extended Kalman filter. The experiments are performed to verify the localization performance of the developed method in several motion scenarios, including a normal operation mode, a wheel slip mode, and a kidnap mode.

This paper presents several different sensor-fusion methods for vision tracking, to keep a camera's line of sight on a target. The data-fusion methods in this paper use several combinations of vision and motion information. The vision information is acquired through a vision processing ASIC (Application-Specific Integrated Circuit), which has an ODR (Output Data Rate) of 27 Hz. The motion information is acquired through one or more sensors, comprising a 3-axis accelerometer, a 3-axis gyroscope, and robot wheel encoders, which have the ODR of 100Hz, 100Hz, and 50 Hz, respectively. Several data-fusion scenarios are tested with these different sensors. An extended kalman filter and an appropriate data-fusion method are applied to each method. Also tested for comparison is a vision-only scenario with a slower ODR of 6.75 Hz. The vision-only method with the fast ODR of 27 Hz performs reasonably well. However, the slower case with the ODR of 6.75 Hz does not provide an adequate performance. The sensor-fusion method that uses all three motion sensors and a vision sensor greatly improves the target tracking performance, regardless of the vision sensor ODR. The motion sensors used are very low cost mass production sensors, found commonly in mobile phones. The proposed method presents a viable solution for target tracking in mobile robots.