Standoff Weapon Imaging Using Mm-Wave Inverse Synthetic Aperture Radar and AI Detection

This research effort seeks to detect large, concealed-on-human, metallic weapons at stand-off distances unobtrusively and inexpensively, using mm-wave radar specifically targeting moving individuals in crowds. By analyzing the reconstructed radar image of pedestrian torsos, gathered for different subject positions as he/she walks adjacent to the radar, it has been shown that it is not sufficient to use the strength or shape of the metallic anomaly reconstruction to distinguish it from the large response of the torso itself. Instead, a new approach has been investigated: imaging, supplemented with artificial intelligence anomaly detection. Artificial intelligence in the form of machine learning is a powerful tool and by learning the common aspects among a set of images, it can distinguish anomalous features and objects in new images. These features are learned by applying multiple convolutional filters in various sizes to different parts of the image [1], [2]. Using convolutional neural networks (CNNs), machine learning architectures have significantly improved the results of image classification. For this project, we are taking advantage of a fully convolutional deep learning model to perform the semantic segmentation task which helps us determine the anomalous responses on a reconstructed radar image and predict the shape of the body-worn gun [3]. Typically, deep learning models need many images to train on. Here, we have generated about almost 2000 simulated radar images using the Finite Difference Frequency Domain (FDFD) computational electromagnetic modeling method. The simulated images include various torso shapes and sizes, for various frequencies, at different angles with respect to the radar. In the simulations, one to three anomalies have been attached to randomly chosen coordinates on subject's skin. There are also multiple simulations in which two torsos (a pair of subjects) are within the computational domain. The reconstructed images have been divided into training, validation, and testing sets and a supervised U-Net model has been trained to predict the anomaly and body. The model can predict the location and size of foreign objects with reasonable accurately. The images are created using a novel, inexpensive multibeam Luneburg Lens/ planar reflector antenna with one mm-wave radar transmitter and three receivers [4]. Multiple views of the subject are recorded as the subject moves from one beam to an adjacent one, to provide high cross-range resolution. Image registration within each radiated beam is cued by simultaneous video trajectory tracking. Several subjects can be tracked and imaged at once.