GestLite: Lightweight Gesture Detection for Edge AI via Event-Driven Model and Input Adaptations

Deep neural networks (DNNs) power many edge vision applications, yet their deployment is often constrained by stringent energy and latency budgets. In this work, we present GestLite, a DNN-based gesture detection system optimized for energy efficiency through joint application–system co-design. GestLite introduces three orthogonal optimization knobs: 1) model complexity, via structured width/depth scaling and head pruning, yielding a family of lightweight YOLO variants; 2) spatial resolution ( $r_{l}$ ), which applies input down-sampling (e.g., $640\rightarrow 160$ px), reducing early-layer MACs by up to approximately 93.75%; and 3) frame stride ( $k$ ), enabling adaptive frame skipping. A lightweight detector runs continuously at low resolution and low frame rate, invoking a high-fidelity model only on gesture-bearing frames. This dual-stage pipeline dynamically adapts its energy footprint to gesture activity through event-driven scheduling. Trained on the HaGRID dataset and deployed on Jetson Orin AGX, Gest Lite achieves up to $3\times $ reduction in average power while maintaining mAP $_{50} \!\gt \! 90\%$ . A full Pareto sweep across 900 configurations reveals that only 2% of designs are Pareto-optimal under joint knob optimization, highlighting the necessity of cross-coupled scheduling. The best configurations deliver F $1~{\gt }97.6\%$ at power budgets