Performance Enhancement of Human Activity Recognition Using Millimeter-Wave Multi-Link Channels

Human activity recognition (HAR) using millimeter-wave (mmWave) radar has gained attention as a contactless and privacy-preserving sensing method that remains effective under low lighting and through non-transparent obstacles. However, conventional monostatic or bistatic deployments with a single viewpoint often struggle in the presence of posture-induced occlusions, as limited sensor placement prevents them from capturing motion cues from alternative perspectives. To address this limitation, we propose a multi-vantage sensing framework that employs two bistatic configurations of mmWave multi-link channel measurements, positioned at the anterior and posterior sides of a human subject. Doppler spectrograms obtained from both perspectives are fused to leverage the diverse motion features captured from each viewpoint. A convolutional neural network (CNN) was trained on the fused spectrograms, and six widely adopted CNN architectures were evaluated to determine the most effective configuration. To further improve classification accuracy, we also developed a hybrid approach that combines CNN-based feature extraction with a support vector machine (SVM) classifier. This configuration yields the best accuracy of 96.79%, confirming the effectiveness of combining deep feature extraction with machine learning classifiers. Experimental results demonstrated that the proposed framework effectively recognizes human activities using multi-vantage sensing based on mmWave multi-link channel measurements.