RRQ-CNN: A Novel CNN for Efficient Human Activity Recognition with Quantization and Ridge Feature Selection

Several state-of-the-art human activity recognition (HAR) studies use convolutional neural networks (CNN) with quantization, where CNN performs the accelerometer-based activity detection. Quantization compresses the model for smart-phone deployment. However, some HAR datasets with many features caused by abundant sensors, axes, and feature extraction types can add excessive complexity to HAR prediction. This study aims to develop and evaluate effective quantization and ridge regression-based feature selection approaches for CNN models (RRQ-CNN) in edge computing-based HAR. We assess the performance of the quantized model in terms of accuracy, inference speed, memory usage, and power consumption on a representative edge device. In this case, we use several metrics such as accuracy, accuracy loss, model size, and compression ratio (CR). This research uses rigorous optimization processes, including a normalized complemented accuracy size curve (NoCASC) for accuracy-efficiency balancing. The test results show that the application of ridge feature selection to CNN HAR performs better in MDI and variance threshold with accuracy values of 0.95, 0.93, and 0.94, respectively. Then, in terms of accuracy, the application of quantization reduces the performance of the basic CNN and the ridge-CNN, with the accuracy losses of 0.186 and 0.01, respectively. What quantization lacks in accuracy makes up for in CR. Quantization provides the two best CRs in the comparison: 4.0× for the quantization+CNN model and 10.0× for the proposed RRQ-CNN. Lastly, through the NoCASC method, we found that RRQ-CNN has a more balanced accuracy-efficiency than ridge+CNN.