Bringing Edge Intelligence to Wildlife Camera Traps with Hyperdimensional Computing

We present a smart wildlife camera trap with an efficient image classification pipeline that can identify high-value animal images while filtering out those triggered by non-animal motion (e.g., moving vegetation and vehicles) or partial animal appearances. To enable low-power, resource-efficient processing, we adopt hyperdimensional computing (HDC) techniques for image analysis. Our evaluation shows that HDC is highly effective for such binary classification tasks. We implement a hardware prototype that is able to autonomously process an image in ~100 milliseconds with 3W max power consumption, enabling the possibility for grid independence and long-term operation. The HDC algorithm achieves more than 88% accuracy on existing datasets and 91% accuracy in a real-time outdoor deployment when identifying animal presence, significantly reducing the number of false-positive image capture events. This paper demonstrates the feasibility of HDC in low-power edge-computing applications like wildlife camera traps.