Research On Night Vehicle Detection Algorithm Based On Deep Learning

At night,the accident rate caused by dark driving conditions is much higher than other time periods.Assisted driving is the key technology to help drivers avoid related accidents,in which vehicle detection is the core module.Traditional methods usually use the single low-level vehicle features for detection,so they have poor robustness in complex and changeable environments.The detection methods based on deep learning can extract higher-level vehicle features with richer semantic information,and they are more adaptable to changing scenarios.However,the existing deep learning methods are usually used in the daytime with good light,and there are many challenges in the night scene: the weak illumination environment makes vehicle features easy to hide;the data set is relatively lacking;the vehicle’s salient features are not fully learned by the network;the single detection method results in only a part of the target’s situation being revealed.All of these make it difficult for existing deep learning methods to show the same detection effect as daytime scenes.In view of the above challenges of the existing methods,a night vehicle detection algorithm based on deep learning is proposed.The main research work is as follows:1.Aiming at the problems that vehicle features are easily hidden and the lack of data sets at night,a preprocessing enhancement algorithm IBIMEF(Improved Bio-Inspired Multi-Exposure Fusion)based on Retinex theory and using multi-exposure images fusion is proposed.In addition,through the study of enhanced algorithms based on different principles,six methods and the IBIMEF algorithm proposed in this paper are analyzed through subjective and objective evaluation indicators.And a data pr ocessing scheme that uses the proposed IBIMEF algorithm for preprocessing and a variety of good enhancement algorithms for data set augmentation is designed.Experimental results show that adding these two operations to Faster R-CNN can improve the average detection accuracy by2.9%.2.Aiming at the problem of insufficient utilization of vehicle saliency information,a salient feature maps generator(SFMG)algorithm is proposed.First,the candidate salient regions such as car lights and reflective car bodies are segmented by adaptive threshold,and then these regions and vehicle labels are used to train the classifier to construct the salient feature map s generator SFMG.The generated salient feature maps are multi-scale transformed by the maximum pooling operation,and finally they are spliced to different convolutional layers of the detection network to be used to fuse with the visual features extracted by the network.The experimental results show that the average accuracy of Faster R-CNN network fused with salient feature maps is improved by 2.7%.3.Aiming at the problem of poor generalization caused by single detection method,a multi-model ensemble algorithm MMEA is proposed.Firstly,based on the ideas of ensemble learning,the algorithm constructs the div ersity of training model from three levels: training data,training strategy and training network types.Then the detection results of multiple models will be used to perform intersection and union operations based on the object overlap ratio and confidenc e.After the operation of removing the redundant detection boxes is performed,the combined results are output as the final results.The MMEA algorithm can take the advantages of different networks to integrate the results of multiple models.Experiments show that the average accuracy of the combined results reaches 86.71%.