AdaptiveStereo: Depth estimation from adaptive structured light

Recovering depth with active stereo methods has become an important approach to three-dimensional perception. However, when the properties of the measurement scene change, projecting predesigned structured light can cause improper exposure and defocus blur on some object surfaces. To address these problems, we propose an active stereo depth perception method called AdaptiveStereo, based on adaptive structured light (ASL). The core idea is to use prior depth information to generate optimal structured light for different measurement scenes, ensuring even illumination of the object’s surface and thereby improving depth perception results. To this end, firstly, a novel ASL encoding scheme based on speckle pattern is designed. Then, we construct an adaptive structured light generation model (ASL-GM) to automatically generate the optimal structured light. Finally, depth perception is realized by using the active stereo matching network with adaptive guidance (AGS-Net). Compared with mainstream depth perception methods, the evaluation results show that our method has higher accuracy, and can obtain better depth perception results. [Display omitted] •Build an adaptive structured light encoding scheme and dataset.•Build an adaptive structured light generation model, addressing overexposure/underexposure and defocus problems.•Propose a depth perception network based on active stereo.