Research On Key Technologies Of Dim And Small Moving Target Detection Under Complex Background

Dim and small moving target detection is widely used in military and civil fields,such as precision guidance and early warning system in military,maritime supervision and space satellite target detection in civil fields.Robust and efficient dim and small moving target detection technology in complex background is the core and key technology in the above fields.And it has always been one of the hot issues in the field of optoelectronics.At present,there is a lack of effective general algorithm for weak and small moving target detection in various complex backgrounds.This paper focuses on the situational awareness scenarios of space satellite targets and the detection scenarios of airborne targets by airborne platforms,and studies the detection of dim and small moving targets under complex backgrounds that are commonly involved.The specific research content includes GEO satellite weak and small target detection in complex geostationary orbit(GEO)satellite map,aerial flying target detection in complex background,and construction of public data set of weak and small moving target in complex background.(1)Aiming at the problem that it is difficult to detect weak and small targets of geostationary satellites in complex GEO satellite images,a weak and small target detection algorithm for GEO Satellite based on image transformation and energy accumulation is proposed.Traditional methods use the difference between target and background characteristics to extract candidate targets on a single frame.Because the target is relatively weak and small,it is easy to miss the detection,and the star line brings great interference to the target detection.This paper uses the stellar stripes in the sequence of images to transform the images between frames,and then accumulates the energy of the weak and small targets based on the dynamic programming method to enhance the signal of the weak and small targets,and finally realizes the reliable detection of the weak and small targets.Compared with traditional methods,the algorithm based on image transformation and energy accumulation has achieved better performance in the aspect of detecting the number of real targets.(2)Aiming at the problem of dim and small aerial target detection under the airborne platform,a weak and small target detection algorithm based on spatio-temporal cues is proposed.Timing information is of great significance for realizing reliable detection of weak and small moving targets.However,the timing information is not fully utilized in traditional methods.In this regard,according to the continuous smoothing characteristics of target trajectory,this paper uses short strict constraint and long loose constraint to parameterize the trajectory,and combines the single frame candidate target extraction based on Light GBM algorithm with the processing of trajectory growth and connection on sequence images to realize the reliable detection of real weak and small moving targets.Compared with existing methods,the target detection algorithm based on spatio-temporal cues has a higher accuracy and recall rate on public data sets.(3)In view of the lack of public image data for dim and small moving targets detection under complex background,this paper constructs the infrared dim and small moving target data set in complex background with the largest scale and the most abundant scene types in the current field,and publishes it publicly(https://small-infraredaerial-target-detection.grand-challenge.org/)to meet the needs of algorithm research and test application.With the help of small UAV platform equipped with infrared detection equipment,through field shooting,data processing and simulation target embedding processing,this paper constructs a set of infrared dim and small moving target data set under complex background with large data scale and rich scene types.Compared with the existing data set,this data set has larger scale and richer scene types,and it is more suitable for algorithm testing and deep learning model training.