Radar Signal Processing for Human Identification by Means of Reservoir Computing Networks

Abstract

Along with substantial advances in the area of image processing and, consequently, video-based surveillance systems, concerns about preserving the privacy of people have also deepened. Therefore, replacing conventional video cameras in surveillance systems with less-intrusive and yet effective alternatives, such as micro-wave radars, is of high interest. The aim of this work is to explore the application of Reservoir Computing Networks (RCNs) to the problem of identifying a limited number of people in an indoor environment, leveraging gait information captured by micro-wave radar measurements. These measurements are done using a commercial low-power linear frequency-modulated continuous-wave (FMCW) radar.Besides the low quality of the outputs of such a radar sensor, walking spontaneously as opposed to controlled situations adds another level of complexity to the targeted use case. In this context, RCNs are interesting tools, given that they have shown a high effectiveness in capturing temporal information and handling noise, while at the same time being easy to setup and train. Using Micro-Doppler features as inputs, we follow a structured procedure towards optimizing the parameters of our RCN-based approach, showing that RCNs have a great potential in processing the noisy features provided by a low-power radar.

Multi-Person Continuous Tracking and Identification from mm-Wave micro-Doppler Signatures

“…2. Our classifier is the most accurate, significantly outperforming the previous DCNN approach [1], the one based on reservoir computing networks (RCN) [10], and performs slightly better than [3], where a structured inference network (SIN) and long-short term memory recurrent neural networks (LSTM) are used. We believe this improvement is achieved due to the use of IBs, which allow for feature extraction at different scales, without significantly increasing the network complexity, which would easily lead to overfitting.…”

“…Human identification from radar sensors is a research theme that is rapidly gaining momentum. Some papers target the classification of the subject identity from the µD signature of gait using radio signals [1], [3], [6]- [10]. Other studies focus on human activity recognition from the backscattered radio signal for security or smart-home applications [5], [11], [12].…”

“…This heavily limits the applicability of the proposed algorithm to real situations, where multiple targets are likely to share and concurrently move in the physical space. The same authors also propose two improvements over their algorithm, to improve its accuracy, but the single-target limitation is still present [3], [10].…”

“…Using the IDRad dataset, we have assessed the performance of our framework for the single person identification problem and have compared it with available benchmarks [1], [3], [10]. For a fair comparison against previous work, we adapted the DCNN to accept as input µD sequences with length of 45 frames instead of 30.…”

Investigating the significance of adversarial attacks and their relation to interpretability for radar-based human activity recognition systems

“…These frames contain velocity and range information about all the objects in the line-of-sight (for the duration of the recording). Recent studies show that with the help of (deep) neural networks, it is possible to leverage these RD and MD frames to recognize multiple individuals [19,47] or to detect human activities with high precision [49]. The aforementioned studies represent these RD and MD frames in the form of a sequence of mono-color images which are supplied as an input to deep CNNs.…”