Study On Intelligent Detection Technology For Multi-source Fusion Target Based On Airborne Platform

For the airborne platform based single sensor,when observing targets,a great challenge is that it is extremely difficult to achieve tasks of data acquisition and information perception under complex environments such like bad weather,thick smog,hard light and other external interferences.By constructing multi-source observation systems to obtain complementary multi-source data information,overcoming the performance degradation or even completely failure of a single sensor under such harsh conditions.A lot of problems,however,are still remain to be urgently solved for multi-source fusion and detection technology.For instance,on one hand,in the face of complex ground scene and significant diverse targets,or different physical mechanisms,spatio-temporal dislocation and large particle size differences of the acquired multi-source data,it is difficult for airborne platform observation systems to obtain multi-source data in sufficient and effective.On the other hand,the existing processing platforms can hardly meet the requirement of the throughput and real-time.Therefore,multisource data real-time processing is a tough nut for multi-source data fusion and target detection.It is extremely important to carry out real-time multi-source data fusion and target detection technology in airborne platform.The main contributions of this work is summarized as follows:1)Aiming at the difficulty caused by the small size of remote target in multi-source photoelectric image data on airborne platform for feature extraction,the original horizontal detection box of photoelectric target detection network method is improved to an arbitrary rotation one;a great loss function is designed to reduce background interferences on small target detection to improve the whole of target features extraction.Compared with the traditional horizontal annotation box method,the experiment results verify the superiority of the proposed method.2)Aiming at the problem of large amount of computation and it is difficult to meet the realtime requirements for the existing multi-source fusion detection method,a decision-level fusion photoelectric target detection method based on multi-source spatio-temporal correlation is proposed.In the proposed method,by using the prior information of the target detection,the spatial and temporal positions of two images,that is,a visible and an infrared image,are matched.Then the decision-level fusion is realized by utilizing the detection results with high confidence in the matched images.The proposed method avoids the huge computation of image pixel-level registration,and realizes real-time photoelectric target detection of airborne platform by decision-level fast fusion.The proposed method has been used in the area of emergency rescue,disaster prevention and mitigation scenarios.The multi-source airborne flight experiments are carried out and the photoelectric image measured results are obtained based on the pod of optic,infrared and laser,verifying the effectiveness of the proposed method.3)Compared with photoelectric sensors,Synthetic Aperture Radar(SAR)has the advantages of long-distance and all weather observation.However,the performance of existing target detection methods based on SAR images remain to be improved due to the large difference in frequency band and resolution of SAR systems.Also,the lack of data sets is a big problem.Aiming at the low resolution and sensitive to speckle noise of SAR image problems in multisource system,which leads to the lack of target feature information and difficult feature extraction problems,in this work,the attention mechanism is introduced into the target detection network to optimize the feature extraction performance and improve the target feature extraction ability of SAR images.Experiments on SAR ship detection dataset and high resolution SAR images dataset demonstrate that compared with some other algorithms,the proposed method achieves better detection performance.Aiming at the poor generalization ability of detection model and poor detection effect caused by the lack of airborne platform SAR data set problems,by exploring the use of transfer learning for optimization,we propose a SAR image target detection method based on multi-band data feature transfer.By using multi-band SAR image data with transfer learning,the network model is adjusted by freezing and fine-tuning operations.The processing results of real measured airborne SAR images show that the average detection rate of the multi-band data feature transfer detection methods is better than some other methods.4)Finally,from the perspective of engineering application,we introduce the experimental data acquisition method of airborne platform experimental system and the airborne data target detection part of multi-source data information perception and elements extraction system,respectively.Specially,the experimental data acquisition method of the airborne platform experimental system includes the experimental system composition and the experimental data acquisition process.The multi-source data information perception and factor extraction system includes the overall structure of the system,the system development environment and the system interface.