Photometric Depth Super-Resolution

Patch-Based Uncalibrated Photometric Stereo Under Natural Illumination

This paper presents a photometric stereo method that works with unknown natural illumination without any calibration objects or initial guess of the target shape. To solve this challenging problem, we propose the use of an equivalent directional lighting model for small surface patches consisting of slowly varying normals, and solve each patch up to an arbitrary orthogonal ambiguity. We further build the patch connections by extracting consistent surface normal pairs via spatial overlaps among patches and intensity profiles. Guided by these connections, the local ambiguities are unified to a global orthogonal one through Markov Random Field optimization and rotation averaging. After applying the integrability constraint, our solution contains only a binary ambiguity, which could be easily removed. Experiments using both synthetic and real-world datasets show our method provides even comparable results to calibrated methods.

Combining time of flight and photometric stereo imaging for 3D reconstruction of discontinuous scenes

Time of flight and photometric stereo are two three-dimensional (3D) imaging techniques with complementary properties, where the former can achieve depth accuracy in discontinuous scenes, and the latter can reconstruct surfaces of objects with fine depth details and high spatial resolution. In this work, we demonstrate the surface reconstruction of complex 3D fields with discontinuity between objects by combining the two imaging methods. Using commercial LEDs, a single-photon avalanche diode camera, and a mobile phone device, high resolution of surface reconstruction is achieved with a RMS error of 6% for an object auto-selected from a scene imaged at a distance of 50 cm.

Millimeter-Wave Array Radar-Based Human Gait Recognition Using Multi-Channel Three-Dimensional Convolutional Neural Network

At present, there are two obvious problems in radar-based gait recognition. First, the traditional radar frequency band is difficult to meet the requirements of fine identification with due to its low carrier frequency and limited micro-Doppler resolution. Another significant problem is that radar signal processing is relatively complex, and the existing signal processing algorithms are poor in real-time usability, robustness and universality. This paper focuses on the two basic problems of human gait detection with radar and proposes a human gait classification and recognition method based on millimeter-wave array radar. Based on deep-learning technology, a multi-channel three-dimensional convolution neural network is proposed on the basis of improving the residual network, which completes the classification and recognition of human gait through the hierarchical extraction and fusion of multi-dimensional features. Taking the three-dimensional coordinates, motion speed and intensity of strong scattering points in the process of target motion as network inputs, multi-channel convolution is used to extract motion features, and the classification and recognition of typical daily actions are completed. The experimental results show that we have more than 92.5% recognition accuracy for common gait categories such as jogging and normal walking.