Yifan Xing



Monocular SLAM

Visual Odometry, Simultaneous Localization and Mapping (SLAM) and 3D reconstruction have become increasingly popular these years with the fast development of robotics and mobile vision applications. Particularly, monocular SLAM only requires an optical camera which is cheap, light, and versatile, and hence can easily be put into commodity system. Generally, feature-based and direct methods are two main approaches of Monocular SLAM. We are particularly interested in applying visual odometry/SLAM algorithm in a highly oscillating environment. We are interested in the performance of both feature based and direct based visual odomoetry/SLAM methods in such an environment. The most cited related works of feature based method include PTAM and ORB-SLAM. Since we are dealing with an oscillating environment that requires robust and wide baseline feature matches, we thus implemented a Feature Based Monocular SLAM system on our own for comparison.

To compare and contrast the feature based method with direct ones. We look into great detail the visual odometry method of ERL , which works based on optical flow weighting and optimisation and adjust it based on the oscillating environment scenario. A MATLAB based visualiser was built for this particular method. Furthermore, a visualisation tool to compare and contrast these two approaches in a uniform manner is built in PCL.

System Pipeline of Feature Based Monocular SLAM System:

As shown in the above pipeline, the system of Feature Based Monocular SLAM that we implemented from scratch consists of these following steps: map initialization by eight point algorithm and RANSAC, data association refinement by Perspective-n-Point algorithm, mapping by bundle adjustment, and 3D dense reconstruction with depth measurement.

Feature Matching and Epipolar Constriant Example:

3D dense resconstuction given depth measurement and 3D Mesh Example:

For detailed discussion of these modules as well as discussion of pros and cons of feature based and direct visual odometry methods, please refer to the project poster. In the following section, comparison results of our implementation of a feature based Monocular SLAM, ORB SLAM method and ERL method are shown on different data sets. Note that for our Monocular SLAM and ORB SLAM, a 3D reconstruction would be available whereas in ERL which is a pure Visual Odometry approach, only the camera pose and trajectory will be shown.

1. Test Result on Hopping Robot Dataset
2. Test Result on Hopping Robot Dataset
Check out the repository on GitHub!
Full Video Demos