Design of a Portable Odometry System using Stereo Vision -- 2
Budget: $30 – $250 CAD
design of a portable visual odometry system that utilizes a feature-based
stereo visual odometry, with IMU assistance. The focus is on developing an algorithm that can provide a usable level of accuracy, while still being possible to implement in a compact and ergonomic fashion. The system uses rectified and distortion-corrected images, and implements
a Harris-Stevens corner detection algorithm to detect landmarks. These landmarks are matched in successive frames using a KLT point-tracking algorithm. Stereo correspondences are found using an optimized block-matching algorithm using the normalized cross-correlation
metric. These disparity values are used to triangulate two-point clouds at successive steps in frames, after which a weighted least-squares pose estimation algorithm is run to estimate the pose change between timesteps. A multivariate gaussian anomaly detection algorithm is used to reject outliers due to poor matching or moving objects.
Finally, a Kalman filter is exploited to enforce a constant-acceleration dynamic model on the cameras, and fuse IMU data for better attitude estimation.
stereo visual odometry, with IMU assistance. The focus is on developing an algorithm that can provide a usable level of accuracy, while still being possible to implement in a compact and ergonomic fashion. The system uses rectified and distortion-corrected images, and implements
a Harris-Stevens corner detection algorithm to detect landmarks. These landmarks are matched in successive frames using a KLT point-tracking algorithm. Stereo correspondences are found using an optimized block-matching algorithm using the normalized cross-correlation
metric. These disparity values are used to triangulate two-point clouds at successive steps in frames, after which a weighted least-squares pose estimation algorithm is run to estimate the pose change between timesteps. A multivariate gaussian anomaly detection algorithm is used to reject outliers due to poor matching or moving objects.
Finally, a Kalman filter is exploited to enforce a constant-acceleration dynamic model on the cameras, and fuse IMU data for better attitude estimation.