Underwater Depth Map Estimation from Video Sequence with Graph Cuts

Low-cost and small-scale underwater vehicles (UVs) are receiving increasing attention due to their small size and superior maneuverability. However, performing autonomous navigation remains a major challenge for small-scale UVs. Conventional sensor suites are unsuitable for small-scale UVs due to size limitations. Thus, cameras are good alternative. Among all the information cameras can offer, depth is one of the valuable sources which can be used for low level obstacle detection and collision avoidance. Though there is intensive research on depth estimation in general, limited literature about depth map estimation for the underwater scenes exist. In this paper, we first use unsupervised color correction method to enhance the underwater image sequences. Then, we estimate camera positions from motion based on feature extraction and matching. Similar to stereo depth estimation, we tackle video sequence depth map estimation as an energy-minimization problem. Lastly, multi-label graph cuts is applied to solve it. The proposed method was tested on a sequence of video and the effectiveness is demonstrated.