This paper explores effective deep learning methods for sensor-based Human Activity Recognition (HAR), emphasizing their implementation in resource-limited wearable devices where Microcontroller Units (MCUs) impose restrictions on memory and processing capabilities. We aim to minimize both computational and memory requirements while ensuring high recognition accuracy. We offer a benchmark comparison of pruning and quantization optimization techniques versus lightweight models enhanced through attention mechanisms and knowledge distillation, from both recognition success and resource-efficiency angles. We evaluate two leading deep learning architectures, DeepConvLSTM and SqueezeNet, across four benchmark HAR datasets: Opportunity, Sensors, Wisdm, and Pamap2. For devices with limited memory capacity, we recommend using lightweight models that integrate attention mechanisms and knowledge distillation. We emphasize that quantization should be prioritized to enhance efficiency, with pruning acting as a secondary approach. Additionally, we provide practical guidelines for deploying optimized HAR models on resource-constrained wearable devices.