A Memory and Computation Efficient Local Stereo Algorithm Suited for Mobile Robot Applications

Abstract

Range sensing is a key component for a mobile robot to interact with the environment. This paper proposes a memory and computation efficient local stereo algorithm for mobile robot applications. The memory reduction is done by enforcing local smoothness constraint during the weighted cost aggregation to make the difference between neighboring aggregated matching cost small and then applying predictive coding scheme to compress the original representations. In addition, by performing cost aggregation only at the sampled positions, the run time is largely reduced and the memory requirement is further decreased. Finally, a high resolution disparity map is derived by an efficient disparity upsampling algorithm utilizing approximated geodesic distance. The tradeoff between matching accuracy, memory cost and computational complexity is extensively investigated using Kitti dataset. Experimental results demonstrate that our method provides high-quality disparity maps with low memory cost and computational effort.