Efficient Transformer-based Model for Human Proximity Detection

Collaborative robotics have the potential to advance the new era of industry by optimizing systems through effective human-robot interaction. Reliable, real-time human detection is essential in this context. In this letter, the authors present a computationally-efficient transformer-based deep learning model for human detection using capacitive-based sensors. Our approach involves a two-stage process: (i) a reconstruction pretext task, where an autoencoder is trained solely on human data, and (ii) a one-class classification stage, as a calibration step, where the frozen encoder acts as a feature extractor to differentiate human from non-human samples. This two-stage pipeline allows rapid calibration in real-world scenarios without further deep learning fine-tuning. Our encoder contains less than 30,000 trainable parameters, and the architecture is modular, allowing for flexible integration of different sensor quantities. We demonstrate the model's robustness through two ablation studies: (i) achieving over 95.44% accuracy with configurations of 1, 2, and 6 sensors, and (ii) achieving over 94.22% accuracy with only 1.4 seconds of recorded human data in a shifting scenario. Additionally, the pipeline attains sub-millisecond inference times on both GPU and CPU devices, underscoring its suitability for real-time applications in collaborative robotics.