Research On Deep Models For Object Detection In Hazy Weather

With the development of social economy,the number of motor vehicles in China is increasing year by year,and more and more people are participating in the transportation.The rich travel demand has also promoted the vigorous development of autonomous driving technology.However,some special traffic scenes pose difficulties for autonomous driving tasks.First,in driving scenes with bad weather,such as hazy and low-light scenes,camera imaging will be greatly disturbed.This scenario not only reduces the performance of the detection algorithm,but also causes false detections and missed detections.Second,the image and video enhancement algorithms in hazy weather are mostly dehazing algorithms based on atmospheric scattering models.The effect of dehazing algorithms on object detection is not clear and the computational cost is high.Third,there are diversity and differences in hazy scenes.The influence of fog on the detection effect of different scene is quite different,and it is impossible to cover all detection problems of fog scenes through one detection scheme.Therefore,it is of research significance to design an effective object detection algorithm in hazy scenes.Aiming to the above problems,this article puts forward the following research points:1.Aiming at the problem of poor robustness of object detection algorithms to hazy scenes,this paper constructs a object detection model based on domain adaptation and attention mechanism.In the feature extraction network part,the model separates and fuses multi-scale convolution feature maps with channel dimensions to extract more refined multi-scale features.In the domain classifier module,the attention mechanism based on global pooling is used to fuse high-level attention to strengthen the domain discrimination ability.Experimental results show that this method can improve the detection accuracy in hazy scenes,and the actual detection effect is better.2.In view of the problems that the current hazy image preprocessing algorithm cannot significantly improve the object detection effect,and the detection-oriented preprocessing model is too complex,computationally expensive,and has many false detections,this paper constructs a detection-oriented,based on switchable normalization image enhancement model.This method combined instance normalization and batch normalization in the model,and fully considered the indispensable role of the two in CNN.Compared with the representative hazy image enhancement methods in recent years,this method has fewer model parameters.The processing efficiency is high and it helps to reduce the occurrence of false detections.3.Aiming at the different applicability of image enhancement and object detection methods in different hazy scenes,this paper further comprehensively evaluates and analyzes the current representative methods based on quantitative detection accuracy and qualitative detection effects.The experimental results show that the application of the attention domain adaptive model proposed in this paper can detects small targets and serious overlaps when directly detecting hazy images.The cascade scheme that first applies the switchable normalization enhancement model proposed in this paper and then performs detection is more suitable for extreme weather scenes with dense fog and low light.