People counting Internet of Things (IoT), which plays a role in counting people indoors based on sensor values, is a vital part of smart buildings because it affects other IoT systems that regulate devices like lighting and air conditioning (AC), impacting efficiency. However, a lightweight solution is needed to perform people counting without threatening personal privacy. This study aims to develop an edge computing-based people counting system using environmental sensors and a Deep Neural Network (DNN) model optimized using the LRF technique. The system is designed to operate in real-time on edge devices with low latency and efficient resource consumption. In general, the system's work process is divided into three main stages, namely (1) data acquisition and pre-processing, (2) model development and optimization, and (3) overall system performance evaluation. The system runs automatically on edge devices and follows a cyclic workflow to detect the number of people continuously. This study also uses ant colony optimization (ACO) for hyperparameter tuning and obtains optimum hyperparameters. Experimental results support the claim that LRF significantly reduces model size while maintaining high prediction accuracy. ACO on hyperparameter tuning obtains the optimum hyperparameters: the number of neurons as many as 128 units, Adam learning rate of 0.005, and batch size of 8. Then DNN + ACO is proven to perform better than DNN without ACO and the state-of-the-art random forest model with accuracy, precision, recall, and F1-score of 0.98, 0.99, 0.94, and 0.97. This is while overcoming the imbalance problem in the dataset with recall for counts 0, 1, 2, and 3, of 1.00, 1.00, 1.00, and 0.78, respectively. This study provides three main contribution. First, this study introduces a novel integration between Ant Colony Optimization (ACO), Deep Neural Network (DNN), and Low-Rank Factorization (LRF) for an edge computing-based people counting system using environmental sensors. Second, the develop system is able to maintain user privacy because it does not use visual data, and improves computational efficiency through model compression using LRF. Third, the proposed model is proven to outperform baseline approaches such as Random Forest and standard DNN, with an increase in F1-Score from 0.93 to 0.97, and a reduction in model size of more than 6% without sacrificing accuracy.
Keywords: Deep Neural Network, Internet of Things, Low-Rank Factorization, People Counting, Ant Colony Optimization