Doppler‐vision

Abstract

Chinese researchers propose an m-D signal separation system, capable of distinguishing between different people walking. AlexNet was used to test the system, where it performed well with different group sizes, achieving nearly 90% accuracy for differentiating between 10 people: a considerable improvement over conventional radar systems. Radar m-D is a frequency modulation that can be applied over the main Doppler shift, caused by tiny movements by the target. When the target is a human, due to our bipedal stance, our arms, legs and torsos can give rise to unique m-D signatures, allowing for differentiation between humans and animals, or to distinguish between different activities: for example, running, crawling or walking. Incredibly, m-D is so sensitive it can be used to differentiate between different humans, based on idiosyncrasies in body movement. This sensitivity makes m-D visualisation suitable for surveillance use in closed-circuit television cameras or other visual sensors. Authors Tao Shan (left) Xingshuai Qiao (centre) and Ran Tao (right). Spectrograms of two human subjects used in the study. Experimental scene used to record movement of each subject Shan et al. identified that there was a dearth of research around using m-D to distinguish between different people doing the same task or activity, especially with regards to large (∼10 individuals) groups of people. They used the deep convolutional neural network (DCNN), one of the most successful deep learning algorithms for image recognition, to analyse limb movements of targets in order to differentiate them. Use of limb movements presents its own set of problems, however, as the radar cross section of a torso is much larger than that of a limb. The elaborate features associated with limb movement are obscured by echoes from the much bulkier torso, meaning that in order to use limb movement as an indicator, the authors needed to develop a method of improving the resolution of the limb's m-D component. Fortunately for the authors, they had already developed a method of separating signals from m-D radar, using a short-time fractional Fourier transform (STFrFT). Two STFrFTs were used with different orders and window lengths to sparsely characterise echoes from limbs and torsos. This is then combined with the so-called morphological component analysis (MCA), a method allowing separation of features contained in an image when these features present different morphological aspects. This combination results in an optimisation problem which, when solved, allows successful separation of echoes from limbs and torsos. The authors used 10 volunteers of mixed gender and varying heights and weights in their experiment. The volunteers were asked to walk normally in front of the radar equipment and were scanned, separating the signals from limbs and torso as outlined previously. The data was fed into a DCNN, AlexNet, and trained. The DCNN performed well at identifying individuals, reaching a success rate of almost 90% for distinguishing between 10 people, 91.25% for 8 individuals, 92.67% for 6 and 97.5% for 4. The high accuracy, say the authors, can be primarily attributed to the separation of limb and torso signals.