Real-time Stereo Matching And Striped Structured Light Depth Estimation

Stereo matching is a crucial step in binocular stereo vision.Stripe structured light depth estimation is an important branch in structured light methods.Both are significant noncontact 3D imaging technologies.As depth learning techniques achieve increasingly high accuracy in both fields,the issues of slow inference speed and large memory consumption become prominent.The speed of algorithms is critical for downstream applications such as autonomous driving,robotic navigation,and industrial inspection.Timely and accurate acquisition of environmental depth information enables relevant devices to make correct decisions and control.This thesis studies real-time stereo matching and stripe structured light depth estimation,with the main work as follows:1.Real-time stereo matching network.This thesis addresses the slow inference speed of current models and the concentration of disparity prediction errors at object boundaries.A bilateral grid upsampling module is designed to improve the inference speed and edgearea disparity estimation.The bilateral grid,which fuses object detail information,can effectively enhance model inference speed while retaining more object details.To optimize edge-area disparity estimation,an edge-based loss function is proposed for model training.Extensive experiments on the Scene Flow,KITTI 2012,KITTI 2015,and Middlebury 2014 datasets demonstrate the simplicity and effectiveness of the proposed method.With a 33 fps(frames per second)condition,the end-point error of the proposed model is as low as 0.63.Moreover,the edge-based loss function can be easily embedded into existing stereo matching networks,such as Gwc Net,AANet,and PSMNet.Embedding experiments show that the edge-based loss function can reduce the end-point error by 3.5%,11.6%,and 27.2%,respectively.2.Stripe structured light depth estimation.Depth estimation of stripe structured light under deep learning has emerged in recent years.To address the scarcity of open-source datasets and sample numbers in this field,we create a large-scale stripe structured light simulation dataset using the open-source Blender software.A simple and effective depth estimation network is designed to achieve fast and accurate depth estimation,and a method using 3D convolution to increase the receptive field is proposed.Extensive experiments on real and simulated datasets show that the 3D convolution feature fusion module can effectively improve depth estimation accuracy by approximately 13.9%;compared to other models,the proposed stripe structured light depth estimation model can achieve higher accuracy at an inference speed of 125 fps;pre-training the model on the simulated dataset results in an accuracy improvement of approximately 4.3% on the real dataset.