Continuous Human Action Recognition by Multiple-Object-Detection-Based FMCW Radar

A continuous human activity recognition method based on the multiobject recognition (MOR) method, the constructed lightweight network (LNet), and the proposed one-dimensional bounding loss (ODBL) function, the MOR LNet ODBL (MOR-LNOD) method, is proposed. The method is validated using continuous action sequences involving nine participants and eight different actions. We interpret each action in the sequence as a single target and utilize a multiobject detection method for accurate single-action region selection, followed by recognition and classification. Based on the results of the study, the MOR-LNOD method is 96.5% accurate on average, which is an improvement of about 20% compared with previous methods based on recurrent neural networks. Compared to the ResNet 50 and MobileNet used in the traditional faster region-based convolutional neural network, the proposed network architecture has reduced the parameters by ten and two times, respectively. As compared to state of the art (SOTA) on the publicly available dataset, MOR-LNOD not only reduces the requirement of input data but also has a higher average accuracy than SOTA.