Comparitive Transforms in Visual Odometry
For Path Localization with Autonomous Vehicles

A comparison of transformation reconstruction methods for optimal path
estimates using Feature Detection, Description, Matching and Trajectory Generation

Mike Knerr

This project estimates an autonomous vehicle trajectory using a sequence of images taken with a monocular camera mounted on a vehicle. The images are from a CARLA autonomous driving simulator that include both RGB and grayscale versions. Depth maps for each image are also utilized.

Reconstruction of the transform between consecutive images consisting of a rotation matrix and translation vector necessary for the trajectory point calculations is made using three different methods: EMD (Essential Matrix Decomposition), PnP, (Perspective-n-Point), and PnP-RANSAC.

A comparison of the trajectory results using EMD, PnP, and PnP-RANSAC with both distance filtered and unflitered feature matches between images is done with visual plots of the trajectories.

The interactive study and tutorial covers these areas:

• Explore the image set, depth maps, and intrinsic camera parameters
• Compare RGB and grayscale images
• Preprocess images with gaussian filtering
• Extract features from the sequence of images
• Match features between consecutive images
• Create distance filtered and unfiltered matched pair sets for calculations
• Use matched feature pairs to estimate camera motion in 3D camera coordinate frame between images
• Build the trajectory paths using three transformation types: EMD, PnP, and PnP-RANSAC
• Plot out results for EMD, PnP, and PnP-RANSAC for both filtered and unflitered feature matching
• Compare results to find the best method
• Observations

_{Project is partially derived and motivated by a suggested continued research direction from the University Of Tornoto course Visual Perception for Self-Driving Cars}