Research On Vehicle Detection Algorithm In Foggy Weather

The booming development of artificial intelligence has prompted computers to replace humans in a lot of work,such as vehicle detection,speech recognition,face recognition and so on.In foggy weather,urban traffic is seriously affected,and vehicle detection is easy to appear wrong detection and missed detection.In order to solve this problem,this thesis relies on the National Key Research and Development Program of China(2019YFE0108300),the National Natural Science Foundation of China(61302150)and other projects to study the vehicle detection algorithm in foggy weather.First of all,due to the lack of vehicle image data set in foggy days,the collected images of non-foggy vehicles are fogged according to natural principles.Secondly,a new defogging model is constructed to defogging the data set.Finally,according to the characteristics of the data set,an algorithm model suitable for vehicle detection in foggy weather is proposed.The main contributions are as follows:1.Due to the existing defogging algorithm’s incomplete defogging effect and the general dimming of the restored fog-free image,a multi-scale fusion image defogging model based on deep learning was proposed.In foggy weather,the vision of human and mechanical equipment is disturbed,so it is necessary to defog the fog-containing data set before vehicle detection.The dehazing model proposed in this thesis uses convolution kernels of different scales to extract multi-scale features from the feature map,and referenced by the ideas of residual structure combines deep and shallow,and then get more abundant information characteristic figure,this approach can effectively solve the problems of the disappearance of gradients and the loss of image details in fog,and the dehazing effect is more thorough.2.In foggy weather,the detection accuracy of common vehicle detection algorithms is generally low.Aiming at the characteristics of the foggy vehicle data set,a model YOLOv4-FOG suitable for vehicle detection in foggy weather is proposed.In this thesis,the trapezoidal feature fusion method is proposed,and the original feature extraction network is simplified and improved,and the trapezoidal feature extraction structure T-CSPDark Net33 is obtained.In addition,since most of the vehicle targets in the data set are small,the prediction layer with the larger receptive field is discarded in the multi-scale prediction stage,and dense connection modules are added to the prediction layer with the small receptive field.Finally,a complete YOLOv4-FOG fog vehicle detection model is obtained.After experimental comparison,the model not only improves the detection speed,but also improves the detection accuracy of vehicle targets in foggy weather,and is suitable for vehicle detection in foggy weather.The m AP value of YOLOv4-FOG reached 79.4%,and the speed reached 41.6FPS,which is 3.2%and 16.9FPS higher than the original model,and has significant advantages compared with other models.3.Due to the lack of traffic vehicle data set in foggy weather,the model training sample is insufficient,so a foggy vehicle data set is constructed.In this thesis,the collected images of non-foggy vehicles are fogged according to natural principles,so as to form sufficient samples.The original fog-free samples are mainly derived from the Stanford CARS data set and traffic vehicle images taken in real environment.The appearance of fog hinders human vision and affects the normal operation of society.Therefore,the research on vehicle detection algorithm in fog weather plays an irreplaceable positive role in the research field of target detection and social development,and also provides an important basis for the relevant departments to make traffic management policies in time.