Toggle Main Menu Toggle Search

Open Access padlockePrints

Insect neural networks as a visual collision detection mechanism in automotive situations

Lookup NU author(s): Dr Richard Stafford, Dr Claire RindORCiD


Full text for this publication is not currently held within this repository. Alternative links are provided below where available.


The lobula giant movement detector (LGMD) of locusts is a large visual interneuron known to respond to objects, such as predators, on a direct collision course. Previous computer models of the neuron interfaced with camera equipped robots have demonstrated its effectiveness at collision avoidance in simple environments. The data presented here examines the performance of a computer realised LGMD model to real and simulated automotive scenes from the perspective of a camera situated inside a car. The model was challenged with videos of a range of non collision traffic sequences as well as videos of collisions with cars. Finally the model was interfaced with a driving simulator game on a Sony Playstation (Sony Corporation, Japan) to assess its performance over a range of collision and non collision sequences. Although the model can detect the majority of the collisions prior to occurrence it falsely signals collisions when vehicles were found to pass in front of the camera from left to right or vice versa or when the car in which the camera was situated turned sharply. False detection of non collision stimuli is not apparent in the real LGMD of the locust and was largely caused due to the differences in speed to size ratio between collisions with cars and approaching avian predators. The LGMD model was then combined with a simple model of the Elementary Movement Detector (EMD) neurons in the fly visual system. These EMD neurons were used to detect translating objects or rotational movement and when triggered inhibited the response of the LGMD. Although the incorporation of the EMD neurons slightly reduced the success rate of detecting collisions by occasionally inhibiting the LGMD when real collisions were occurring, the number of false detections was substantially reduced. The study shows that the mechanisms of visual processing used by insect neurons can be exploited for commercial purposes. However, because the commercial uses of the neuron do not exactly match the purpose for which it evolved in the insect, modifications to the neurons or integrating several neurons which process information in different ways may be required for robust performance.

Publication metadata

Author(s): Stafford R, Keil MS, Yue S, Cuadri J, Rind FC

Publication type: Conference Proceedings (inc. Abstract)

Publication status: Published

Conference Name: Insect Sensors and Robotics

Year of Conference: 2004