Radar-based Human Activity Recognition (HAR) offers a privacy-preserving alternative to camera systems for Ambient Assisted Living (ADL), particularly in scenarios where visual monitoring is inappropriate. This paper proposes a modular deep learning pipeline for ADL recognition using short-range 60 GHz FMCW radar. The system combines Doppler-based signal encoding, feature extraction, comparing 16 pre-trained convolutional neural networks, and temporal modeling using attention-enhanced bidirectional LSTMs. To overcome the challenge of manual ground-truth annotation—a major bottleneck in HAR dataset creation—an automated labeling framework based on RGB-D and pose estimation is introduced, enabling efficient training data generation in continuous or unscripted sessions. Five representative ADLs (dress, sit, stand, walk, lying down) were recorded in a generic indoor setting using a compact radar sensor. Performance was evaluated across all model configurations, with InceptionV3 achieving the lowest misclassification error rate of 0.019 and average classification accuracy approaching 98%. MobileNetV2 also delivered high accuracy with reduced complexity, confirming its suitability for deployment on low-power edge/embedded platforms. The proposed pipeline demonstrates competitive performance and generalizability, offering a robust foundation for scalable, radar-based HAR in privacy-sensitive environments.