Dilated Causal Convolution Based Human Activity Recognition Using Voxelized Point Cloud Radar Data

Due to the immense advantages that include contactless sensing, privacy-preserving, and lighting condition in-sensitivity, radar systems have been applied in Human Activity Recognition (HAR). The radar signal is often used in its raw form, pre-processed into micro-Doppler signatures or represented as voxelized Point clouds. However, the point cloud data is usually sparse and non-uniform. HAR deep learning models ought to learn the spatial and temporal features. These models should be robust for all considered activities and computationally efficient. Instead of other deep learning techniques used in literature, dilated causal convolutions (DCC) provide a broad receptive field with a few layers while preserving the resolution of the inputs throughout the model, thereby learning the spatial and temporal cues. In this paper, we investigated the use of DCC in combination with other deep learning techniques like residual blocks (RDCC), transformer encoders (TED), and bidirectional long-short-term memory (BiLSTM). We subsequently proposed the DCCB model that consists of DCC layers and BiLSTM layers. The proposed model exhibits a commendable performance in terms of accuracy, and generalization especially in terms of balanced robustness for all activities.