Feature Based Landmark Extraction for Real Time Visual SLAM

Recently, there has been a marked increase in using machine learning techniques for object detection because of its immunity to noise and variations in backgrounds. SLAM (Simultaneous Localization and Mapping) is an approach to mapping the environment in which the robot moves, by using landmarks, much like the human visual system. The application of a robust object detection system can be extended into the field of SLAM by using these as powerful visual landmarks. While the traditional approach to SLAM (based on sensors like the SONARs or LASERs) can provide a good perception of depth, they cannot form effective landmarks. The output of these devices contain the range data mapped on a 2D space. The landmark has to be significant to show up as a pattern and hence only significant landmarks get extracted. While the visual information may be more than enough to form very good landmarks, the required computational resource increases way beyond the realm of the present day embedded processors. We use GPUs (Graphic Processing Units) to process the visual information since they have been very successful in doing real-time rendering for graphics application which involve similar mathematics. The presence of a large number of cores makes this a challenging problem to solve as programming them can be quite complex to exploit the full bandwidth of these processors. Much work is going on to integrate these units into embedded devices which make it feasible to solve the problem of visual SLAM.