Research On Key Technologies Of Vehicle Detection In Haze And Rain

With the rapid development of the field of artificial intelligence and the advancement of deep learning theory,vehicle detection technology has been widely applied in the field of urban intelligent transportation and autonomous driving.It assists vehicles to achieve safe driving in complex driving environments,while helping regulators monitor vehicle behavior,improve traffic flow and increase traffic safety.Under good weather conditions and visual quality conditions,vehicle detection technology has achieved a high level.But under challenging typical severe weather conditions(rain or haze)and under the influence of large-scale small objects,image quality degradation causes detection not effectively.Therefore,this paper studies how to overcome the impact of bad weather conditions on image data quality and ensure the accuracy of the detection system.It is of great significance to improve the safety of autonomous driving and the reliability of intelligent transportation systems.In this thesis,for the image data collected by the visible light camera,a series of problems such as image quality degradation and low vehicle detection accuracy under typical severe weather conditions in hazy and rainy days are studied.It focuses on image quality enhancement,image dehaze and derain,and vehicle detection algorithms.The work and innovations of this paper are mainly in the following four aspects:(1)An improved generative adversarial network image super-resolution reconstruction method for small object-oriented vehicle detection is proposed.In complex traffic scenes,there is a problem that small objects at a long distance occupy fewer pixels,leading to a significant decrease in vehicle detection accuracy.In this paper,an image super-resolution object detection algorithm is designed based on Transformer.First,the linear embedding module of Transformer is improved according to the structure of Generative Adversarial Network to supplement the linear embedding information of small objects.At the same time,the edge enhancement network is introduced to enhance the edge of the object,which improves the resolution of long-distance small objects in the image,dense parking areas on both sides of the road,and vehicle objects under occlusion.Then,the acquired high-resolution images are hierarchically constructed,similar to the feature pyramid in convolutional neural network,each layer only models the local relationship,and uses the shift window mechanism to expand the receptive field,replacing convolutional layer to extract more powerful small object vehicle features,to achieve accurate detection.Through comparative experiments,it is proved that the algorithm proposed in this paper improves the detection performance of vehicles in the case of small scale and serious occlusion,and provides a basis for the subsequent research of vehicle detection algorithms in severe weather.(2)A vehicle detection algorithm based on improved U-shaped network in hazy conditions is proposed.On the one hand,aiming at the difficulties in feature extraction caused by low quality,blurred objects and serious loss of detailed features in hazy images affected by particle media in the air,an end-to-end dehazing network is designed with Transformer as reference to the structure of U-Net.The learning ability of the network is improved by introducing a cross-attention skip connection mechanism to fuse multi-scale features,recovering images with clean features after dehazing.On the other hand,aiming at the difficulty of vehicle detection in hazy images,a vehicle detection model for hazy conditions is constructed by combining dehazing and detection.Experiments on synthetic and real hazy image datasets show that the algorithm can obtain clear dehazing images and effectively improve the accuracy of vehicle detection in hazy scenes.(3)A multi-branch feature fusion vehicle detection algorithm under rainy conditions is proposed.Aiming at the problem of low vehicle detection accuracy in rainy days,a Transformer based multi-branch structure derain model is designed.By extracting different levels and different types of features,the model’s ability to represent complex rain patterns is improved,so as to effectively remove rain patterns and recover the lost information of degraded images.The detection model combines the advantages of Transformer and convolutional neural network to design a local perception enhanced Transformer backbone network,which enhances the local perception ability of the algorithm.At the same time,in order to solve the problem of few image resources in rainy days,this paper creates a rainy dataset for traffic scenes.Experiments prove that the algorithm can improve the vehicle detection performance in the corresponding scene while enhancing the image quality of rain degradation.(4)A vehicle detection algorithm under rain and haze conditions based on multilevel feature fusion is proposed.In view of the fact that haze and rain often occur at the same time in the real scene,a model is designed to remove haze and rain at the same time without switching between a single model and training data.In this model,a hierarchical encoder is used to obtain multi-level features of degraded images,that is,multi-scale features with different resolutions,to better capture the useful information for image dehazing and deraining,and restore the details of the degraded image.At the same time,a decoder is introduced that can learn the weather type query.The embedding of this weather type learns with the network to adapt to the same weather type as the input image.Finally,combined with the image dehazing and deraining model,the overall framework of the vehicle object detection model with local perception enhancement is proposed,and the model is tested using mixed weather conditions and single weather condition datasets.The experimental results show that the algorithm has good image recovery effect and high detection accuracy in mixed rain and haze weather,which meets the needs of practical applications.