OppLoD: the Opponency based Looming Detector, Model Extension of Looming Sensitivity from LGMD to LPLC2

Feng Shuang, Yanpeng Zhu, Yupeng Xie, Lei Zhao, Quansheng Xie, Jiannan Zhao, Shigang YuearXiv 2023 · DOI · arXiv

Introduction

For autonomous systems such as unmanned aerial vehicles (UAVs) and mobile robots, "looming detection"—the ability to identify an object on a direct collision course—is a fundamental requirement for survival and navigation. While biological systems have evolved elegant solutions to this problem, translating these into robust computational models remains a challenge. This paper introduces OppLoD, a hybrid bio-inspired neural network that merges the strengths of two distinct insect visual systems: the locust and the fruit fly. By combining these mechanisms, the authors have developed a model that is significantly more sensitive and robust than previous state-of-the-art detectors.

The Challenge of Looming Detection

Most existing bio-inspired collision models rely on the Lobula Giant Movement Detector (LGMD) neuron found in locusts. The LGMD is highly effective at tracking image velocity and detecting expanding edges. However, simple expansion detection has a critical flaw: it can be easily confused by "ego-motion" (the robot's own movement) or general background clutter, leading to false positives.

To address this, the authors look to the LPLC2 neuron in Drosophila (fruit flies). Unlike the LGMD, the LPLC2 is "ultra-selective" for Radial Opponent-Motion (ROM). This specific pattern occurs when motion vectors point outward from a central point, creating a symmetric "explosion" of pixels. By focusing on this radial symmetry, the system can distinguish a true looming threat from environmental noise or translational movement.

Biological inspiration from the Drosophila LPLC2 neuron and the concept of motion symmetry.

OppLoD Architecture

The OppLoD model extends the existing Distributed LGMD (D-LGMD) framework by integrating several novel modules designed to mimic the fly's LPLC2 neuropile:

  1. Motion Direction Extracting (MDE) Module: Standard models often use uniform filters. MDE instead uses anisotropic kernels to encode specific motion directions (up-right, down-left, etc.), mimicking the four layers of synaptic fans found in the fly's visual system.
  2. Opposing Motion Judgement (OMJ) Module: This module calculates the correlation between opposite motion directions using lateral inhibition and a self-attention mechanism. It confirms if an object is truly looming by verifying the symmetry of the outward motion.
  3. Zoned Receptive Fields: To handle objects that are not perfectly centered, the model zones the receptive field. This allows it to detect radial symmetry locally, significantly improving performance for off-center obstacles.
Comparison between traditional expansion-based detection (a) and the proposed OppLoD pipeline (b).

Experimental Validation and Results

The model was evaluated using both synthetic image sequences and real-world indoor video datasets featuring approaching objects like balls and cubes, some mounted on remote-controlled cars to simulate unpredictable movement.

  • Enhanced Sensitivity: In synthetic tests, OppLoD reached its peak response at frame 22, whereas the baseline D-LGMD model did not peak until frame 43. This provides a critical time buffer for evasive maneuvers.
  • Robustness to Noise: The model successfully ignored inward motions and non-symmetric expansions that typically trigger false alarms in simpler models.
  • Real-World Performance: In dynamic tests, OppLoD began sensing threats much earlier than baseline models and maintained a significantly higher signal-to-noise ratio, demonstrating its readiness for complex environments.

Conclusion and Future Work

OppLoD demonstrates that computational synergy between different biological species can yield superior robotic vision systems. By combining the locust's velocity sensitivity with the fruit fly's radial selectivity, the model achieves a new level of "ultra-selective" looming detection. Future work will focus on integrating specialized neurons to eliminate shadow interference and validating the model in high-speed outdoor UAV scenarios where ego-motion is more aggressive.

Summary updated

BibTeX
@misc{shuang2023opplod,
  author = {Feng Shuang and Yanpeng Zhu and Yupeng Xie and Lei Zhao and Quansheng Xie and Jiannan Zhao and Shigang Yue},
  title = {{OppLoD: the Opponency based Looming Detector, Model Extension of Looming Sensitivity from LGMD to LPLC2}},
  year = {2023},
  eprint = {2302.10284},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2302.10284},
}
Figure