Structured light-based shape measurement system

It is very important to acquire accurate depth information of target object or scene for many applications in machine learning. The use of 3D reconstruction based on active laser triangulation technology is very complex in real application. One main problem is that most of these technologies detect light stripes by considering each column or row of the image as independent signals causing lack of robustness. In real application, variable illumination, uneven surface and imaging noise will make stripe detection fail. In this paper, by considering laser stripe distortion assumption, we adopt efficient belief propagation algorithm to extract center of laser stripe, which proves superior to existing peak detection approaches. Because of occlusion and low reflectivity, laser stripe captured by the sensor will be cut into several parts at some points, which are referred to as outliers. As for non-outlier, the SNR of that point is high and the disparity difference between left and right neighbor is slight. First, determine whether a point is an outlier or not by computing the weighted SNR and disparity difference. Then efficient belief propagation algorithm is adopted to infer the outlier map which is called labels in machine learning. Experimental results demonstrate the feasibility of our proposed approach.

► Proposed an efficient shape measurement approach based on structured light. ► Adopt efficient belief propagation algorithm to extract center of laser stripe. ► Detect outliers according to SNR of laser stripe. ► Proposed an approach to deal with occlusion.