Efficient Hand Detection with Low-resolution Infrared Sensors

In the evolving landscape of Machine Learning (ML) and Deep Learning (DL), there is an increasing emphasis on developing models that efficiently balance performance and computational constraints, rather than focusing exclusively on accuracy. This work investigates efficient ML and DL models for detecting the presence of hands and other objects using low-resolution infrared (IR) sensors, which represent a cost-effective, low-power, and privacy preserving alternative to higher resolution cameras. To support this study, a specialized dataset was collected, containing images of hands, objects, hands with objects, and empty backgrounds. The dataset includes a variety of objects with different sizes and temperatures, providing a comprehensive foundation for robust model training and evaluation across diverse detection scenarios. The implications of this research study extend to various domains, including robotics, industries, healthcare, and sign language recognition. The study begins with a comprehensive analysis of traditional ML techniques, specifically Support Vector Machines (SVM) and Random Forest (RF), to establish baseline performance metrics. Subsequently, the research explores the capabilities of DL models, in particular Convolutional Neural Networks (CNNs), known for their ability to capture intricate patterns in the visual domain. This exploration follows a two-phase approach. The first phase deals with the design of a CNN with the aim of maximizing the sole detection accuracy. The second phase considers balancing accuracy with model complexity, addressing the need for models that can operate effectively in resource constrained scenarios, such as edge devices. To achieve these objectives, the study integrates Bayesian Optimization with Differentiable Neural Architecture Search (DNAS). Specifically, the OpTuna framework is combined with two main gradient-based techniques: Pruning In Time (PIT) and Mixed-Precision Search (MPS). This hybrid strategy allows for precise architectural tuning, ensuring the models achieve a balance of high detection accuracy and optimal resource usage, adaptable to various performance and computational needs. Experimental results indicate that traditional ML techniques, such as RF and SVM, achieve accuracy rates of 80% and 85%, respectively, while CNNs outperform them with an accuracy of 90%. These findings underscore the potential of CNNs in applications requiring higher precision, although traditional methods, particularly RF, offer advantages in computational speed. The optimization phase demonstrates that it is possible to significantly reduce model size with limited impact on prediction capabilities.