Efficient multi-model integration neural network framework for nighttime vehicle detection

A vehicle detection system is a core ADAS function for automatic driving. However, owing to the low-light environment, nighttime vehicle detection is a big challenge. Current techniques have the limitation of not being able to fully extract the nighttime vehicle features. To solve the problem, this study proposes a nighttime framework, which employs multiple means to enhance nighttime vehicle information. First, multiple image enhancement techniques are used to rich the training datasets, and an improved Bio-Inspired Multi-Exposure Fusion (BIMEF) algorithm is proposed to improve the quality of the nighttime images. Then, the multi-scale salient feature maps of highlight vehicle regions are combined with vehicle visual feature maps to enhance vehicle information during detection. At last, an ensemble algorithm is proposed to combine multiple networks to provide richer vehicle visual features. Experiment results show the effectiveness of our method in terms of accuracy and speed. Additionally, our method is robust in multiple complex night scenes.