A Disparity Fusion Method for a Dual-Baseline Multi-Camera Stereo Vision System

To address the difficulties of conventional binocular stereo vision in recognizing some weak-texture regions and in simultaneously ensuring ranging accuracy and result completeness in scenes with large depth variations, a dual-baseline multi-camera stereo vision platform is developed. The platform consists of a left-right branch and a left-top branch, which share the same left camera and generate two disparity maps with different baseline directions. A multi-disparity fusion method is then proposed for the dual-baseline disparity fusion problem from two aspects: registration and fusion. First, an analytical registration method is used to achieve spatial alignment of the dual-baseline disparity results. Second, a multi-disparity fusion model is constructed by exploiting the advantages of the two baselines at different distances. The experimental results show that the analytical registration method significantly increases the edge overlap rate and reduces the mean edge distance of the dual-baseline disparity results, thereby effectively improving the spatial consistency of the disparity maps. The proposed multi-disparity fusion model enables the stereo vision platform to achieve ranging errors generally below 0.5% within 2 m, below 2% within 4 m, and below 5% within 12 m, indicating a high level of ranging accuracy. In addition, the fused results effectively improve disparity completeness in some weak-texture and occluded scenes.