| With the increasing requirements for the active safety and intelligence of automobiles,the automatic driving performance and object detection methods in various weather conditions have also attracted more attention.However,images captured in foggy conditions show large color contrast,poor saturation,and the presence of targets such as cars,pedestrians,and motorcycles being obscured by haze,which will seriously affect the target detection performance.At the same time,at this stage,the number of labeled object detection datasets in foggy weather scenarios is extremely small.Therefore,based on the domain adaptive target detection network,this thesis improves and uses a foggy target detection algorithm based on domain adaptation.The main work of this thesis is as follows:(1)Aiming at the problem that the domain adaptive object detection method is applied to the problem that the time domain shift is particularly serious when sunny day data is used as the source domain and foggy days as the target domain,this thesis uses a foggy target detection algorithm based on attention mechanism to introduce the attention mechanism module.The attention mechanism can be used to give higher weight to the areas and channels that need to focus on feature extraction,so that the network can focus its attention on the domain invariant features.In order to further improve the strength of domain adaptive targeting,this thesis designs the attention mechanism of light and dark channels,in order to improve the accuracy of domain adaptive target detection in cross-weather domain target detection.(2)In view of the problem that the real scene is often covered by haze in the foggy target detection task,especially for the neural network,it is difficult to extract effective target information through the haze,this thesis introduces the dark channel dehazing algorithm to dehaze the foggy image preprocessing.At the same time,based on the dark channel dehazing algorithm,the feature map scattering transformation is designed and embedded in each stage of Res Net50 to further reduce the influence of haze on target detection,in order to improve the accuracy of target detection in the foggy sky image that can cross weather scenes and be carried out by domain adaptive target detection.(3)In view of the problem that the gradient degradation of the domain classification module is much faster than that of the more complex object detection module in the actual training process of the multi-task model domain adaptive object detection network composed of the object detection module and the domain classification module,resulting in the domain classification module being unable to continuously provide the domain invariant features for the object detection module,the multi-task loss optimization method is introduced to automatically assign a weight to each module loss,so as to balance the convergence speed of different modules in order to achieve better detection results.Experimental results show that the improved and used foggy target detection algorithm based on domain adaptation,which refers to Cityscapes source domain data training and Foggy Cityscapes target domain data to detect eight target categories in traffic scenes,has an average accuracy of 45.1%,which is very competitive compared with other domain adaptive object detection algorithms. |
