Before deep networks, robust systems thrived on carefully engineered features. Band‑pass filters, FFT bins, delta energies, and temporal grammars still shine when power budgets are tight. We share reference numbers, data collection tricks, and annotation habits that reduce confusion, including balanced classes and adversarial examples like jacket sleeves imitating swipes on rainy, hurried evenings.
Before deep networks, robust systems thrived on carefully engineered features. Band‑pass filters, FFT bins, delta energies, and temporal grammars still shine when power budgets are tight. We share reference numbers, data collection tricks, and annotation habits that reduce confusion, including balanced classes and adversarial examples like jacket sleeves imitating swipes on rainy, hurried evenings.
Before deep networks, robust systems thrived on carefully engineered features. Band‑pass filters, FFT bins, delta energies, and temporal grammars still shine when power budgets are tight. We share reference numbers, data collection tricks, and annotation habits that reduce confusion, including balanced classes and adversarial examples like jacket sleeves imitating swipes on rainy, hurried evenings.
Battery life shapes feasibility. We estimate current draw for idle presence scans, bursty gesture sampling, inference, and radios. Techniques include duty cycling, wake‑on‑motion, and compressive sensing. For mains power, we consider electromagnetic interference and grounding, ensuring nearby dimmers, refrigerators, and chargers will not pollute readings or create mysterious, intermittent resets under peak loads.
Where you mount matters more than marketing claims. Corners produce echoes, glossy tiles reflect beams, and fabric absorbs sound. We map coverage with quick tape tests and cardboard mockups, revealing blind spots near shelving, stairwells, and sinks, and showing why eye‑level angles often outperform ceilings unless careful beam shaping and masks are applied.