SIFT Based Graphical SLAM on a Packbot

We present an implementation of Simultaneous Localization and Mapping (SLAM) that uses infrared (IR) camera images collected at 10 Hz from a Packbot robot. The Packbot has a number of challenging characteristics with regard to vision based SLAM. The robot travels on tracks which causes the odometry to be poor especially while turning. The IMU is of relatively low quality as well making the drift in the motion prediction greater than on conventional robots. In addition, the very low placement of the camera and its fixed orientation looking forward is not ideal for estimating motion from the images. Several novel ideas are tested here. Harris corners are extracted from every 5 th frame and used as image features for our SLAM. Scale Invariant Feature Transform, SIFT, descriptors are formed from each of these. These are used to match image features over these 5 frame intervals. Lucas-Kanade tracking is done to find corresponding pixels in the frames between the SIFT frames. This allows a substantial computational savings over doing SIFT matching every frame. The epipolar constraints between all these matches that are implied by the dead-reckoning are used to further test the matches and eliminate poor features. Finally, the features are initialized on the map at once using an inverse depth parameterization which eliminates the delay in initialization of the 3D point features.