| The detection and tracking on moving vehicle is great of significance to establishment of the urban safety emergency response system.At present,the urban monitoring system based on ground video system is restricted by vision,which is difficult to meet the needs of largescale vehicle detection and tracking.Video satellite is a new type of earth observation satellite,which can continuously observe a certain area,and record the observation information by video.Thus,it is conducive to the monitoring of moving objects.Traditional moving object detection methods are prone to be disturbed by background noise,which are highly depended upon the empirical experience to decrease the efficiency of the detection.Therefore,it is necessary to study on the intelligent detection methods of moving objects that is suitable for satellite video.The deep learning technology has great potential in the image processing due to its advantages in automatical learning abilities.This study selects a typical moving object,e.g.,vehicle,as the research object,to explore moving vehicle detection methods by using satellite video data and deep learning technology.Due to the limitation of the spatial resolution regarding satellite video images,the size of vehicles in the video is constrained to a dozen pixels,and the appearance features(color,texture and other features)of the vehicles may be omitted.However,satellite video has the advantage of high time resolution.Most of the existing moving vehicle detection methods in satellite video based on deep learning do not make full application of the temporal information from satellite video,which affects the detection effect.To address the above issues,the main contents of this study is given as follows:(1)A moving vehicle detection dataset in satellite video was created.This study used Labelimg to label the moving vehicles in satellite video,and created Sky Sat moving vehicle detection dataset and Jilin-1 moving vehicle detection dataset,which meet the data requirements for high-quality moving vehicle detection.(2)The intelligent detection method of moving vehicles by taking spatial-temporal information was studied.Since the moving vehicle detection methods in satellite video based on deep learning do not make full use of the spatial-temporal information,this study improved the detection addressing the features of high spatial-temporal resolution of satellite video.YOLOv5 was selected as the detection method,by introducing the Focal Loss function,to improve the feature extraction ability.The concept of frame difference was introduced,to expand the possibility of YOLOv5 model by changing the form of single frame input to the continuous frame triple input.The results show that:(1)The proposed method for construction of the moving vehicle detection dataset is applicable;(2)The expansion of the original data by augmentation methods and the introduction of Focal Loss function can improve the detection ability;(3)By introducing the frame difference module into the YOLOv5 model,the features of high time resolution of satellite video are highlighted to improve the detection ability compared with the conventional deep learning models.By combining the frame difference module with the conventional data augmentation methods and Mosaic data augmentation method,the spatialtemporal information is further enhanced to increase the detection accuracy,which is adaptable to a more complex environment.In summary,by adding the frame difference module to the YOLOv5 model,the features of high time resolution are embedded to improve the detection ability.By combining the frame difference module with the conventional data augmentation methods and Mosaic data augmentation method,the model detection can be improved to adapt to a more complex environment.The proposed method by taking spatial-temporal information has a good effect in the moving vehicles detection of satellite video,which is expected to apply to the scenario related to urban safety emergency response. |
PDF Full Text Request