Long-Range Non-Line-of-Sight Imaging Based on Projected Images from Multiple Light Fields

Abstract

Non-line-of-sight (NLOS) imaging technology has shown potential in several applications, such as intelligent driving, warfare and reconnaissance, medical diagnosis, and disaster rescue. However, most NLOS imaging systems are expensive and have a limited detection range, which hinders their utility in real-world scenarios. To address these limitations, we designed an NLOS imaging system, which is capable of long-range data acquisition. We also introduce an NLOS object imaging method based on deep learning, which makes use of long-range projected images from different light fields to reconstruct hidden objects. The method learns the mapping relationships of projected images and objects and corrects the image structure to suppress the generation of artifacts in order to improve the reconstruction quality. The results show that the proposed method produces fewer artifacts in reconstructions, which are close to human subjective perception. Furthermore, NLOS targets can be reconstructed even if the distance between the detection device and the intermediate surface exceeds 50 m.

Study of depth NLOS imaging by TOF camera using deep learning

“…[6][7][8][9] In recent years, with the continuous development of deep learning technology, the use of highly nonlinear neural networks to solve scattering model problems has attracted widespread attention. [10][11][12] XiaoJie et al 11 proposed a U-Net deep neural network to achieve imaging of NLOS targets 50m away. Dayu et al 12 used VAE deep neural networks to generate NLOS targets under conditional control.…”