Non-Line-of-Sight Target Recognition Based on Multi-Detection Points Collaboration Strategy

Non-line-of-sight (NLOS) target recognition, which refers to the identification of objects beyond the line of sight, like around the corner, has widespread applications in unmanned driving, counter-terrorism operations and many other fields. Current research in this field mainly focuses on passive detecting and identifying speckle patterns generated by hidden objects, of which the feasibility has been proved. However, the performances in practical scenarios, like the accuracy and speed, are far from satisfactory. In this work, we rely on the active NLOS detection system and propose a multi-detection points collaboration strategy to achieve the recognition of NLOS targets. By taking advantage of the high spatial detection efficiency of the single-photon avalanche diode (SPAD) array and by getting rid of the redundant information hidden in the SPAD-array measurement, the hidden target can be recognized fast and accurately. The preprocessing steps for the original recorded histogram are designed to eliminate the unwanted elements, form the collaborated histogram and then enter a designed convolutional neural network obtaining the classification result. Specially, a number of key variables, which are mutually constrained in selecting the pixels for recognition, are discussed, including the number of pixels in one group used to form a collaborated histogram, the spacing between any two pixels in one group, and the total number of groups injected into the network in one measurement. The proposed strategy and network are validated by experiments, which are conducted on both single hidden object and multiple hidden objects in one scene, and achieve nearly 92% recognition accuracy. As the accuracy, speed and robustness of the network are guaranteed, it shows great potential for NLOS target recognition in practical scenes.