Study On Intelligent Detection And Fusion Method Of Moving Target Based On Multi-source Feature Transfer

A single sensor cannot achieve high-reliability and high-precision continuous observation due to weather,light,and interference when observing the target scene.In order to overcome the impact of severe weather conditions on the observation performance of a single sensor,multi-source sensors are used to complement each other's advantages.However,in actual application scenarios,there are still many problems in the complementary technology of multi-source sensors.In the field of disaster search and rescue,the complex survey environment at the disaster site will make the observation data obtained by multi-source sensors more complicated.Obtaining key target information from the redundant multisource observation data takes a long time.There are problems of high computational cost and poor real-time performance for target detection.In the field of continuous monitoring,SAR sensors are used to complement the limitations of photoelectric sensors.The transfer learning effect of a traditional single source domain is greatly affected by the similarity between the source domain and the target domain.And the key parameters also rely on empirical settings.There is a problem of unstable performance.Therefore,it is very important to carry out the studies of multi-source moving target detection technology.The breakthrough of its key technology is of great significance to improve disaster search and rescue capabilities in the context of complex ground objects and realize all-weather highreliability observation of the target scene.Facing the need for rapid detection of multi-source target information in the field of disaster search and rescue,this thesis studies the use of multi-source photoelectric sensor data for target detection.In order to solve many existing problems such as large amount of calculation and poor reliability in multi-source sensors use data layer fusion for target detection,this thesis designs a multi-source data feature joint network for multi-source photoelectric sensor result layer fusion.First,use self-generated sample data to train the designed multi-source data feature joint network,Then use the target detection algorithm based on deep learning theory to perform target detection on the visible light and infrared images respectively.Finally,the correction parameters output by the multi-source data feature joint network are used to fuse the photoelectric target detection results.Compared with the data layer fusion method need a large number of SIFT feature point selection,de-duplication and registration calculations,the result layer fusion calculation cost is relatively low and the calculation time is not affected by the image size.It can meet the real-time requirements of multi-source photoelectric target detection in disaster search and rescue.The photoelectric image data recorded on the pedestrian bridge and the photoelectric image of fire scene personnel search and rescue are used as experimental data to verify the effectiveness of the designed network.Facing the demand for all-weather high-reliability observation of target scenes in the field of continuous monitoring,this thesis carried out studies on target detection with multi-source SAR image data.The traditional single source domain transfer learning can expand the auxiliary training data set of the model.But the migration effect is greatly affected by the similarity between the source domain and the target domain.And the level parameters of the model weight "freeze-adjust" also rely on empirical settings.Aiming at the problem of unstable transfer learning effect of a single source domain,this thesis proposes a SAR image target detection method based on multi-source feature transfer.First,the TP-CFAR algorithm with high false alarm parameters is used to perform preliminary target detection on SAR images to obtain preliminary detection results with a higher false alarm rate.Then use the visible light image data and infrared image data as the source domain to perform migration learning to obtain the corresponding migration model.Calculate the accuracy and migration weights of the two models to obtain the preliminary multi-source feature model weights.Then use the labeled SAR image data set to fine-tune the weights of the multi-source feature transfer model to obtain the multi-source feature transfer model.Finally,the multi-source feature transfer model is used to eliminate false alarms from the preliminary detection results of SAR images to obtain the target detection results.The algorithm proposed in this thesis improves the accuracy of detection through the transfer learning of SAR image multi-source domains.The SAR image of Piraeus Port in Greece was used as measurement data to verify the effectiveness of the multi-source domain migration algorithm.Finally,this thesis introduces the realization process of photoelectric image target detection based on multi-source features and realization process of SAR image target detection based on multi-source feature transfer from the perspective of engineering application.Including the overall structure of the system,the software and hardware environment developed,the actual completion interface and the complete use process.