Monocular SLAM
The research in monocular SLAM technology is mainly based on the EKF(Extended Kalman Filter) SLAM approaches.
Basic Concepts
Extended Kalman Filter
Many SLAM system use EKF to fuse information from different types of sensors such as sonar or laser ranging. For vision based SLAM, using EKF to dealing with data fusion from video sequence is also a popular method. The steps of one iteration of Extended Kalman Filter is given below. For more information, please take a look at the EKF site at CMU.
Extended Kalman Filter based Monocular SLAM
A typical EKF based monocular SLAM system Scenelib is first given by Andrew. J. Davison in the paper " MonoSLAM: Real-Time Single Camera SLAM". Currently, a whole new RT-SLAM system is put forward by Joan Solà etc., which can dealing multiple sensor fusion including vision in real-time. For most vision based SLAM systems, the camera state is comprised of position, rotation quaternion, velocity and angular velocity of the camera, and the fature state is represented by a 3D world coordinates.
Camera State:
Feature State:
EKF System State:
For the EKF predicted step, a motion model is used to get one step prediction of the state. For vision only system, we donot have any kinematic information, so a constent velocity model is used as below.
If some kinds of kinematic sensor such as IMU(Inertial Measurement Unit) is used, the accelerate motion model is always the best way.
For most vision based system, a pinhole model is used to represent camera measurement.
The next graph demostrates the flowchart of a typical monocular SLAM system.
IMU-Camera Calibration
There are something more should be take good care of, the alignment of the coordinates frames of camera and IMU. These stuff can be calibrated through classic hand-eye calibration algorithm.
Some Works
Accelero meter & Multi-level Data Association
|
A real-time monocular SLAM method with the acceleration information from an IMU(Inertial Measurement Unit) is proposed within the framework of Extended Kalman Filter(EKF). In this method, the acceleration information outputted from an IMU is introduced to the filter's process model for reducing the accumulated system error as well as the drift of the state vector.Simultaneously, a corner-based hierarchical data association algorithm is introduced to reduce the uncertainty of data association and the likelihood of mismatch between the observation vectors in the filter's observation model and the feature vectors in the map. (Demo) |
Using IMU Roll & Pitch
|
A monocular SLAM method based only on Roll and Pitch output from an IMU is proposed. We observed that among the three output Euler angles from a low-cost IMU, Roll and Pitch are more accurate than Yaw, then a calibration algorithm for the SLAM system using only Roll and Pitch is presented. With the calibrated results, an IMU-vision SLAM system is implemented robustly, where an EKF is employed to estimate the state vector composed of the camera location, Yaw of the IMU and the locations of the map features. Experimental results show that our proposed method can reduce the accumulated errors effectively and alleviate the negative in°uence of the IMU's measurement errors on the system's stability. (Demo)
|
Real-Time Relocalization
|
A real-time relocalization algorithm based on a hierarchical bipartite graph model is proposed for recovering an SLAM system automatically after tracking failures. When the SLAM system is lost, a hierarchical bipartite graph model is introduced for finding sufficiently good correspondences between the detected image features and the stored map features, and then the effecient real-time relocalization is achieved with the obtained correspondences. The proposed model takes both temporal and spatial constraints on these features into account, which increases the robustness of the relocalization signifficantly. (Demo) |
Some AR(Augmented Reality) Tests
A lovely girl animation on the desk.