Active Non-line-of-sight Imaging Modeling And Reconstruction Based On Phong Lighting Model

Due to the obstruction,conventional imaging technologies can’t obtain the obstructed scene information from the camera’s direct line of sight.Non-line-of-sight(NLOS)imaging techniques typically scan a visible relay surface with a pulsed laser,which scatters light to the NLOS scene,and then capture the time-of-flight information returned by the three scattering with an ultra-high time-resolved detector,aiming to reconstruct hidden objects.As an emerging technology,with the development of detection equipments and the improvement of related algorithms,NLOS imaging technologies are expected to be applied to real life and play a critical role in autonomous driving,remote sensing,endoscopy,reconnaissance and anti-terrorism.Most existing active NLOS imaging algorithms assume that light isotropically scatters on the relay surface,while ignore the actual specular reflection.Under this assumption,NLOS imaging is transformed into solving an ill-posed nonlinear inverse transmission model,which requires complex computation and a large amount of memory.In addition,the high cost of obtaining experimental data and the weaker signal acquired hinder the development of NLOS imaging.In order to solve the above problems,this dissertation constructs a more perfect NLOS imaging model,designs a transient data generator to quickly obtain the transient simulation data through relay walls with the different reflectance properties,providing the data support for the research of NLOS imaging.And then the data-driven high-precision NLOS reconstruction is realized.The main research work include:To address the neglect of specular reflection in existing imaging models,the confocal NLOS imaging model based on Phong lighting model is constructed.Firstly,the scattering process of the laser on the relay surface is analyzed.Based on the confocal scanning device and the Phong lighting model,a light transmission model considering both the diffuse contributions and the specular contributions of the relay surface is constructed.Secondly,in order to simulate the sampling process of single photon avalanche diode(SPAD)more realistically,Gaussian filter is used to simulate the jitter of the detector,and additive noise is introduced to simulate the dark count of SPAD and the background noise.Furthermore,the noise model of active NLOS imaging is constructed,combined with the nature that photons conform to non-uniform Poisson distribution.To solve the problem of small number and high acquisition cost of real transient images,a transient data generator based on Phong lighting model is designed.In this paper,the confocal NLOS imaging model based on Phong model is discretized,and a transient data generator is designed to simulate the transient images in different experimental scenes.Then,the synthesized transient data is analyzed and imaged using the classical light-cone transform,by which the feasibility of the designed generator is verified.In addition,the influence of specular reflection on the physical imaging model is analyzed qualitatively and quantitatively by changing the relevant parameters of the generator,demonstrating that with the increase of the shininess factor,that is,the smoother the relay surface is,the more serious the distortion of the reconstruction by the conventional algorithm is.Focusing on the presence of the specular reflection on the relay surface and more accurate NLOS reconstruction,a end-to-end neural network called Res-UNet with 3D encoding and 2D decoding is proposed.Combining UNet network and residual network,the three-dimensional encoder is used to effectively extract the spatial-temporal information of the transient data.Two-dimensional decoder is utilized to quickly reconstruct the geometric shape and depth of hidden objects,and a regression network composed of two fully connected layers is introduced to further improve the accuracy of depth images.The experimental results demonstrate that the mean square error of Res-UNet imaging is about one tenth of that of two widely used algorithms in the presence of specular reflection.