研究目的
To develop a computational periscopy technique that enables non-line-of-sight imaging using only an ordinary digital camera, without the need for expensive, specialized equipment or controlled illumination, for applications such as monitoring hazardous environments and detecting hidden objects.
研究成果
The study demonstrates that 2D color non-line-of-sight imaging is feasible using an ordinary digital camera by leveraging the penumbra of an occluding object. The method accurately estimates the occluder position and reconstructs hidden scenes with moderate resolution, showing robustness to noise and lack of prior knowledge. This approach reduces the cost and complexity compared to existing methods, with potential applications in surveillance and hazardous environment monitoring. Future work could focus on improving computational efficiency and handling more complex scenarios.
研究不足
The technique requires the occluding object to have a known size and shape, and it may not perform well with highly complex or unknown occluders. Computational complexity is high, with initial occluder estimation taking up to 18 minutes. The method assumes Lambertian reflection and may be sensitive to model mismatches, such as non-diffuse surfaces or ambient light variations. Resolution is limited by the camera and downsampling, with smaller features being less accurately resolved.
1：Experimental Design and Method Selection:
The method involves capturing a single photograph with an ordinary digital camera of a diffusely reflecting surface (imaging wall) that has a penumbra cast by an occluding object. The recovery is based on computational inversion of the linear dependence of the penumbra on the hidden scene, modeled through ray optics and the rendering equation. Algorithms for occluder position estimation and scene reconstruction using total variation regularization are employed.
2：Sample Selection and Data Sources:
The hidden scene is displayed on an LCD monitor for controlled testing. The occluding object is a black rectangular object of known size and shape, but its position is unknown. Additional experiments use 2D and 3D diffuse reflecting scenes as described in supplementary information.
3：List of Experimental Equipment and Materials:
A 4-megapixel digital camera (FLIR Grasshopper3 model GS3-U3-41S4C-C), a 20-inch LCD monitor (Dell model 2001FP), a black rectangular occluding object (7.7 cm × 7.5 cm), a white Lambertian surface (Elmer's foam board), a Tamron M118FM16 lens, and a laptop computer (Lenovo ThinkPad P51s) for control and processing.
4：7 cm × 5 cm), a white Lambertian surface (Elmer's foam board), a Tamron M118FM16 lens, and a laptop computer (Lenovo ThinkPad P51s) for control and processing.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The camera captures multiple exposures (e.g., 20 exposures of 175 ms each) of the imaging wall, which are averaged to reduce noise. The raw image is processed by de-interleaving color channels, averaging green channels, and downsampling to 126×126 pixels per channel. The occluder position is estimated via grid search, and the hidden scene is reconstructed using optimization algorithms with TV regularization.
5：Data Analysis Methods:
Data analysis involves solving linear systems and optimization problems (e.g., using the fast iterative shrinkage-thresholding algorithm) to estimate the occluder position and reconstruct the scene. Spatial differencing and ensemble methods are used to improve robustness and reduce noise.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
