Multi-domains based human activity classification in radar
Z. Li, Francesco Fioranelli, Shufan Yang, Lei Zhang, Olivier Romain, Qian He, Guangyan Cui, Julien Le Kernec
Abstract
In human activity recognition (HAR) based on radar, significant research exists on statistical features extracted from the spectrogram (μD), whereas the research which considers other domains is less developed. This paper is aimed to investigate three domains of radar data: μD, Cadence Velocity Diagram (CVD), and range-time (RT) information, evaluating which ones are best suited to classify specific activities. In addition, information fusion is applied to enhance classification accuracy and compare it with the results of single domain approach. Based on the previous results, a hierarchical structure is proposed to improve the performance of classification further. The preliminary results show that different domains have distinctive sensitivity to specific activities. RT information is sensitive to the moving target crossing range bins, while CVD is more sensitive to body movement. The μD is more balanced, which means it can observe both moving targets and body movements. Furthermore, improvement in accuracy is approximately 6-23 % using feature-level fusion. A hierarchical classification approach is also investigated, which has accuracy in the order of approximately 92 %.