Low-power neuromorphic virtual tripwire for drone detection at ECCV 2024

Drone Detection Using a Low-Power Neuromorphic Virtual Tripwire (ECCV 2024 Workshops) proposes an event-camera-based detection architecture designed for continuous, low-power operation: a virtual tripwire that triggers an alarm whenever a drone crosses a monitored line-of-sight without requiring a person watching the sensor output.

The virtual tripwire concept

A physical tripwire or laser curtain covers a single plane. A virtual tripwire defined in software can be placed at any distance and orientation the sensor’s field of view covers, and can be reconfigured dynamically. Combined with a sensor that monitors continuously, a virtual tripwire provides perimeter coverage without physical infrastructure.

The challenge is running this detection continuously at low power. A conventional frame camera processing system might consume 10–50 watts for continuous detection, which requires grid power or large battery packs. At those power levels, a sensor can monitor for hours but not for days.

The paper targets sub-100 mW total system power, which brings the battery-life requirement down to a practical range for deployed sensors running from primary batteries.

The neuromorphic implementation

The key design choice is to use an event camera’s native motion detection rather than running an inference model over frames. A drone crossing a monitored line-of-sight generates a burst of events as it occludes background pixels and reveals new ones at its leading edge. This event burst can be detected by simple threshold logic — count events per unit time in a spatial region, trigger if the count exceeds a threshold — rather than by deep learning inference.

Simple threshold logic is implementable in a microcontroller or programmable logic device at milliwatt power levels. The system architecture is: event camera (sensor) → event counting (microcontroller/FPGA) → threshold comparison → trigger.

The paper demonstrates this architecture on a field-programmable gate array at 78 mW total power, detecting drones crossing a monitoring plane at ranges up to 150 meters.

Performance characteristics

Detection probability at 150 meters for a Group 1 UAS is reported at above 90% with a false alarm rate of less than one per hour under the tested outdoor conditions. At shorter ranges (50 meters), detection probability reaches near 100%.

The performance degrades in rain and heavy wind (background events from vegetation motion increase the false alarm rate) and against small fixed-wing UAS (which generate fewer events per unit time than multirotors of the same size).

Relevance to persistent site surveillance

The low-power operation is the defining characteristic here. A ground-based C-UAS sensor network that monitors a perimeter needs to run continuously at deployed sites that may lack infrastructure power. A sensor that consumes 78 mW versus one that consumes 50 W has fundamentally different logistics — the former can run for weeks on a small battery pack, enabling truly persistent monitoring.

For airborne or interceptor applications, power budget matters for different reasons — not battery life but heat dissipation and payload allocation. The low-power neuromorphic tripwire architecture is not directly applicable to an airborne seeker, but the event-counting detection approach, which avoids running deep learning inference entirely, is a useful architectural pattern for the cueing and handoff stages that precede a seeker engagement.

---

Drone Detection Using a Low-Power Neuromorphic Virtual Tripwire. ECCV 2024 Workshops.