Research And Analysis On Dehazing Algorithm For Traffic Scene Understanding

At present,artificial intelligence technology is gradually infiltrating into all aspects of people’s lives.It has also brought the transformation and upsurge of automotive intelligence in the automotive field,and has become a major development direction of the industry,leading the forward promotion of the industry’s cutting-edge technology.Vehicle environment perception technology based on vision sensor is an important research topic in the field of intelligent vehicle automatic driving,which is easily affected by various extreme weather conditions.Aiming at the traffic scene in hazy days,this paper first improves the deep learning based dehazing network,and then evaluates and analyzes the dehazing effect of typical dehazing network and the applicability of dehazting work in this field combined with the specific application in the field of intelligent vehicle scene understanding.First of all,three representative dehazing algorithms are compared and tested,and the dehazing effects of different dehazing networks are tested on the synthetic test set and the real haze test set.The test results show that the existing dehazing algorithm performs better on the synthetic test set,but has poor adaptability to the real haze image.Secondly,in order to improve the poor adaptability of the existing dehazing network in real hazy images,the dehazing network is improved on the basis of the existing network.The generative adversarial dehazing network is selected as the main structure of the dehazing network,and the multi-scale dense feature fusion module is introduced to reduce the loss of spatial information in the process of image feature extraction.The experimental results show that the improved network has good dehazing effect on SOTS outdoor synthetic haze test set and RTTS4322 real haze test set.Then,aiming at the application of intelligent vehicle scene understanding,the analysis and evaluation method of existing dehazing network is constructed to evaluate its dehazing effect and applicability on real hazy images.Aiming at the two kinds of scene understanding methods of object detection and semantic segmentation for intelligent vehicle environment perception,from the process of test set preparation,dehazing network selection and testing,scene understanding network selection and testing to result analysis,the dehazing evaluation methods are constructed respectively,and the application effect of dehazing in the field of scene understanding is evaluated.Finally,according to the proposed dehazing evaluation method,the experiment of dehazing evaluation is carried out.For the task of target detection and semantic segmentation,7 typical dehazing networks are applied to pre-process the real haze test set images,and the accuracy of target detection and semantic segmentation before and after dehazing is evaluated.Through the analysis of the comparative test results,some suggestions are put forward for the future dehazing research work in the field of intelligent vehicle scene understanding.