AI-based night-time image recognition techniques are applied through the collection, annotation, and training of nighttime image data to enable effective feature extraction. Based on night-time and long-distance vehicle image datasets, the model architecture is improved and parameters are optimized. By refining the model’s feature structure, the transmission and integration of image features across different layers are improved, allowing the model to more accurately capture vehicle contours and light source features in night-time environments. In addition, targeted fine-tuning is performed for challenging scenarios such as long-distance detection, light glare interference, low-light conditions, and small objects, thereby enhancing the model’s recognition performance under low illumination and small-object settings, and enhancing both stability and accuracy in real world nighttime scenarios.

1.Detection range: approximately 200 m, with a width of approximately 24m in front of the vehicle (UN-R 149) 

2.Applicable road curvature radius: ≥ 200 m (ADB final rule) 

3.Detection categories: cars, motorcycles 

4.Avoid misclassification of streetlights and reflective objects 

5.Average image processing time : ≥ 10 fps

This system can be applied in various types of vehicles equipped with image sensing system.

The technology can make system integrators, automotive electronics suppliers, and image processing suppliers have capabilities of system integration. As a result, they will get opportunities of getting into the image sensing system supply chain to meet future trend of emerging market demand for advanced driver assistance systems or automated driving systems.