ODD-SEC: Onboard Drone Detection with a Spinning Event Camera

This paper presents ODD-SEC, a real-time onboard drone detection system for moving carriers that utilizes a spinning event camera and a novel motion-compensation-free representation to achieve 360-degree surveillance with high accuracy under challenging conditions.

The Gist

Imagine you are trying to spot a tiny, fast-moving drone in the sky while you are riding a bumpy, spinning merry-go-round. That is basically the challenge this paper solves.

Here is the story of ODD-SEC, a new system designed to catch drones even when the camera itself is moving and spinning.

1. The Problem: The "Static" Camera Trap

Most security cameras are like statues. They stand still and take pictures (frames) of the world.

• The Issue: If a drone flies fast, a normal camera sees a blur. If the sun is too bright or it's too dark, the camera goes blind.
• The "Event" Camera: Scientists invented a special camera called an Event Camera. Instead of taking full photos, it acts like a super-fast nervous system. It only "feels" when a pixel changes brightness. It's like a room full of people who only shout "Hey!" when something moves. This makes it incredibly fast and great in bright sun or total darkness.
• The Catch: These event cameras usually have a narrow view (like looking through a straw). If a drone flies outside that straw, you miss it. Also, most systems assume the camera is standing still on a tripod. But what if the camera is on a robot dog running around? The movement creates chaos that breaks the software.

2. The Solution: The "Spinning Lighthouse"

The team built a system called ODD-SEC (Onboard Drone Detection with a Spinning Event Camera).

• The Hardware: They took that narrow "straw" camera and mounted it on a motor that spins it around 360 degrees, like a lighthouse beam sweeping the horizon.
• The Result: Instead of a narrow straw, the camera now sees a full panoramic circle around the robot. No matter which way the drone flies, it will eventually cross the camera's path.
• The Challenge: Spinning a camera creates a massive amount of "motion blur" and data chaos. If you just feed this spinning data into a normal AI, it gets confused and thinks the whole world is moving, not just the camera.

3. The Secret Sauce: The "Time-Slice Sandwich"

To fix the spinning chaos, the team invented a clever way to process the data.

• The Analogy: Imagine you are trying to watch a movie, but the projector is spinning wildly. The image is a mess.
• The Fix: Instead of trying to watch the whole spinning mess at once, the system cuts the video into tiny, frozen slices of time (like slicing a loaf of bread).
• The Magic: They feed these slices into a special AI brain (a modified version of a famous object detector called YOLOX). This AI has a special "Time Fusion" module. Think of this module as a conductor in an orchestra. It listens to the different slices, figures out which sounds (pixels) belong to the drone and which are just background noise caused by the spinning, and then harmonizes them to find the target.
• No Motion Compensation: Usually, you have to mathematically "undo" the camera's spin to see clearly. This system is smart enough to learn the patterns of the spin and ignore them automatically, like a surfer who stays balanced on a moving wave without thinking about the wave's physics.

4. The Real-World Test: The Robot Dog

They didn't just test this in a lab. They put the system on a quadruped robot (a robot dog) and ran it outside.

• The Setup: The robot dog ran around, and a DJI drone flew nearby. The camera spun 360 degrees on the dog's back.
• The Conditions: They tested it in blinding sunlight, in the dark, and while the robot was jiggling and turning.
• The Result: The system worked like a charm. It found the drone in real-time (about 22 times per second) and told the robot exactly where the drone was pointing (within 2 degrees of accuracy). That's like spotting a friend in a crowd and pointing your finger at them with almost perfect precision, even while you are running.

Why Does This Matter?

• Security: It allows robots (like police dogs or security bots) to hunt down bad drones without needing a human to hold a camera.
• Robustness: It works when normal cameras fail (too bright, too dark, too fast).
• Mobility: It proves you can have a 360-degree "super-vision" system that works while moving, which is a huge step forward for autonomous robots.

In short: ODD-SEC is a spinning, super-fast eye on a robot dog that never gets dizzy, never gets blinded by the sun, and can spot a tiny drone flying anywhere around it, all while the robot is running a marathon.

Technical Summary

1. Problem Statement

The rapid proliferation of drones necessitates robust detection systems for security and privacy. While traditional frame-based cameras are versatile, they struggle with fast-moving targets, motion blur, and adverse lighting (low light or high glare). Event cameras offer a solution with high dynamic range (HDR) and microsecond-level temporal resolution. However, existing event-based detection systems face two critical limitations:

1. Limited Field of View (FoV): Standard event cameras have narrow FoVs, making it difficult to track unpredictable drone trajectories without constant re-orientation.
2. Static Assumption: Most current event-based algorithms assume a static camera. They fail when deployed on moving carriers (e.g., quadruped robots, UGVs) because the system's own ego-motion causes severe spatial smearing and tracker failures, rendering standard motion compensation techniques ineffective or computationally prohibitive.

2. Methodology

The authors propose ODD-SEC, a real-time onboard system designed to detect drones and estimate their bearing on moving platforms.

A. Hardware Architecture

• Spinning Mechanism: An event camera (DVXplorer) is mounted on a stepper-motor-driven spinning platform. This achieves a 360° horizontal Field of View, overcoming the FoV limitations of static sensors.
• Onboard Integration: The entire sensing and computing unit (including an NVIDIA Jetson Orin NX) spins together with the camera. This eliminates the need for slip rings or wireless data transmission, ensuring stable, high-bandwidth data flow.
• Synchronization: A photoelectric switch defines a "zero-point" direction per rotation, providing a synchronization signal to the camera to ensure accurate angular orientation recovery.

B. Algorithm Design

The system processes raw event data through a pipeline designed to handle continuous rotation and ego-motion without explicit motion compensation:

1. Multi-slice Spatio-temporal Representation (MSR):

   • Instead of accumulating events into a single image (which causes severe blur during rotation), the system divides a fixed time window ($\Delta T$) into $N$ uniform sub-intervals.
   • Events are binned into sequential "slices." This decouples target kinematics from rotation-induced spatial smearing, preserving fine-grained temporal cues.

2. Temporal Fusion Module (TFM) within YOLOX:

   • The detection backbone is a modified YOLOX network.
   • Feature Channel Gating (FCG): An attention mechanism that adaptively weights feature channels to amplify salient target features and suppress background noise caused by ego-motion.
   • Motion Context Encoder: Processes the sequence of slices to encode cross-slice kinematics, generating a Temporal Context Vector (TCV). This allows the network to learn spatial alignment and spatio-temporal correlations implicitly, bypassing the need for optical flow or IMU-based compensation.

3. Scale-Aware Loss (EIoU):

   • To handle the extreme scale variations of drones (due to changing distances and rotation), the standard IoU loss is replaced with Efficient IoU (EIoU). This explicitly decouples penalties for center point distance, width, and height, ensuring robust regression for both small, distant targets and large, close ones.

4. Bearing Estimation:

   • The system calculates the 3D bearing vector relative to the carrier using the drone's 2D centroid, camera intrinsics, and the real-time rotation angle of the spinning platform (derived from the photoelectric switch trigger and motor speed).

3. Key Contributions

1. Novel System Design: The first onboard drone detection system utilizing a spinning event camera to achieve 360° panoramic perception on moving carriers, effectively solving the FoV limitation of traditional event setups.
2. Motion-Robust Algorithm: A novel image-like event representation (MSR) and a Temporal Fusion Module that enable accurate detection without explicit motion compensation, specifically addressing the challenges of massive data volume and motion blur caused by carrier movement.
3. Real-World Validation: Comprehensive outdoor experiments on a moving quadruped robot (Unitree Go2-W) demonstrating reliable detection and bearing estimation in real-time.

4. Experimental Results

The system was evaluated on an NVIDIA Jetson Orin NX (16 GB) in both static and dynamic (moving robot) scenarios under varying lighting conditions (standard, low-light, and high-glare).

• Detection Performance:
  • Achieved an Average Precision (AP) of 62.19% under standard lighting, dropping only slightly to 55.80% in high-glare conditions and 59.30% in low-light.
  • The Temporal Fusion Module improved AP by +2.73% and AP75 by +3.10% compared to a baseline without temporal fusion.
• Bearing Estimation Accuracy:
  • Under moving conditions, the system achieved a mean angular error of approximately 2.0° (specifically 1.91°–2.06° across sequences).
  • The maximum error remained within acceptable bounds (approx. 4.3°), confirming robustness against dynamic disturbances.
• Computational Efficiency:
  • The system operates in real-time at ~22 Hz on the Jetson Orin NX.
  • The inference bottleneck is handled efficiently via TensorRT (FP16), with the total detection node runtime averaging ~45ms.

5. Significance

ODD-SEC represents a significant advancement in anti-drone surveillance and neuromorphic robotics.

• Operational Versatility: By enabling 360° detection on moving platforms, it bridges the gap between laboratory event-camera research and real-world field operations (e.g., perimeter security by patrol robots).
• Robustness: It demonstrates that event cameras can outperform frame-based systems in extreme lighting and high-speed motion scenarios, provided the data representation and network architecture are adapted for continuous motion.
• Open Science: The authors plan to open-source the complete pipeline, fostering further research into dynamic event-based perception.

In summary, ODD-SEC successfully integrates hardware innovation (spinning platform) with algorithmic novelty (motion-robust temporal fusion) to solve the critical challenge of detecting fast-moving, unpredictable targets on dynamic, moving carriers.