A comprehensive list of visual SLAM solutions

MoNoSLAM

Using the extended Kalman filter as the backend, tracking the very sparse feature points on the front end, using the current state of the camera and all landmark points as state quantities, updating its mean and covariance.

Advantages: In 2007, with the improvement of computer performance and the system's sparse way of processing images, this solution enabled the SLAM system to run online. (The previous SLAM system basically could not operate online. It could only rely on the data collected by the robot to carry the camera, and then perform positioning and mapping offline.)

Disadvantages: MoNoSLAM has shortcomings such as narrow application scenarios, limited number of road signs, and easy loss of coefficient feature points. Its development has now been stopped.

PTAM( Parallel Tracking And Mapping )

The main principle is: capture feature points from the photographic image, then detect the plane, establish virtual 3D coordinates on the detected plane, and then synthesize the photographic image and CG. Among them, the unique feature is that the detection of the stereoscopic plane and the synthesis of the image are processed in parallel.

Advantages: It proposes and implements the parallelization of the tracking and mapping process, separates the front and rear ends, and uses a nonlinear optimization scheme, which can not only position and map in real time, but also superimpose objects on the virtual plane.

Disadvantages: The scene is small and tracking is easily lost.

ORB-SLAM (inherit and improve PTAM)

Advantages: Versatility: Supports three modes: monocular, binocular, and RGB-D. The entire system is calculated around ORB features, achieving a balance between efficiency and accuracy, and is optimized around feature points. Its loopback detection algorithm can effectively prevent the accumulation of errors. Using three threads to complete SLAM achieves better tracking and mapping effects, and can ensure the global consistency of trajectories and maps.

Disadvantages: It is time-consuming to calculate ORB features for each image. Three threads bring a greater burden to the CPU, and there are certain difficulties in embedded devices. The mapping of ORB-SLAM uses sparse feature points, which can only meet the positioning function.

LSD-SLAM（Large Scale Direct monocular SLAM）

The monocular is directly applied to the semi-dense monocular SLAM. There is no need to calculate feature points and a dense map can be constructed.

Advantages: The direct method is based on pixels; it is not sensitive to feature missing areas, and semi-dense tracking can ensure the real-time and stability of tracking; it realizes the reconstruction of semi-dense maps on the CPU.

Disadvantages: It is very sensitive to camera internal parameters and exposure, and is easily lost when the camera moves quickly. In the loop detection part, it is not directly implemented based on direct hair. It relies on the feature point equation for loop detection, and it has not completely got rid of the calculation of feature points.

SVO（Semi-direct Visual Odoemtry）

In the implementation of the visual odometry based on the sparse direct method, 4x4 small blocks are used for block matching to estimate the movement of the camera itself.

Advantages: Extremely fast, can achieve real-time performance on low-end computing platforms, and is suitable for situations where computing platforms are limited.

Disadvantages: Poor performance in head-up cameras; the back-end optimization and loop detection parts are abandoned, SVO's pose estimation has cumulative errors, and it is not easy to reposition after loss.

RTAB-MAP (SLAM scheme on RGB-D sensor)

A complete RGB-D SLAM solution is given. Currently, its binary program can be obtained directly from ROS, and its APP can be used directly on Google Project Tango.

Advantages: The principle is simple; supports RGB-D and binocular sensors, and provides real-time positioning and mapping functions.

Disadvantages: High integration, huge size, difficult to carry out secondary development on it, suitable for SLAM application rather than research use.

The above is the detailed content of A comprehensive list of visual SLAM solutions. For more information, please follow other related articles on the PHP Chinese website!