Vision Based Tracking of Multiple Dynamic Obstacles for Autonomous Robot Navigation

In this paper we present a methodology for vision based detection and tracking of dynamic obstacles for real time autonomous robot navigation in dynamic, unstructured and possibly unknown environment. The problem in visual tracking is its computational complexity for real time applications. This limitation is addressed by scaling down the area of interest in the search space in successive frames of images and hence reducing the time complexity. We decompose the problem of tracking into two steps: (1) narrowing visual area into a smaller area of interest, where the probability of finding the obstacle is high and then, (2) searching for the obstacle in this area. The proposed technique uses Kalman Filter for predicting the area of interest in successive frames and extraction of pseudo phase in two dimensional discrete cosine transform (2D-DCT) based motion estimation in the area of interest to localize the obstacle within it. This algorithm has been successfully tested and verified for various obstacles dynamics and the results show that the computational complexity has been considerably reduced.