Research On Sensing Techniques Of Various Forms Of Photoelectric Targets In Airborne Environment

With the rapid development of unmanned platform technology and sensor equipment,the airborne platform has been able to simultaneously equip various detectors such as visible light and infrared light,therefore,more accurate positioning can be achieved.With the characteristics of these detectors,and the flexibility of drones,the multi-mode unmanned airborne electro-optical imaging platform is widely used in the civil fields such as traffic monitoring and field search and rescue.This paper focuses on the target detection,recognition and tracking under the airborne platform.The main research contents and achievements are as follows.(1)Aiming at the low adaptability of weak and small target detection methods in infrared images,a method of infrared image dim target detection based on convolutional neural network is proposed.Firstly,the pre-training of the model is performed by using the existing image sets;and then the network is modified with the idea of the screening operation;finally,using the parameters obtained from training to process infrared images contain weak targets.From the experimental results,the method proposed in this chapter has a good detection effect.(2)Aiming at the problems of the low accuracy of the original YOLO detection method and the slowness of the network training,a fast end-to-end photoelectric target detection method was proposed.First,the original YOLO full-connection layer was removed,and Faster RCNN’s predefined box mechanism was used to accelerate the processing speed of the network;Secondly,through the analysis of the application environment and the target characteristics,the training set,verification set and test set are made separately for the specificity of the detection target;Finally,the shallow convolution features are combined with the deep features and sent to the classifier for network training,which improves the detection effect of the network on smaller-sized targets.By testing a large amount of data,it can be found that the proposed method outperforms the other two network models in terms of accuracy.(3)For the deep learning tracking method,there is a problem that the target cannot be tracked or even lost when the target is occluded.This paper proposes a self-renewing anti-occlusion target tracking method based on multi-layer depth feature.Firstly,the pre-trained network is used to extract the target multi-layered convolution features.Then,the correlation filter is trained and the response map is obtained,the estimated position of target can be obtained in the region where the maximum correlation response value is located.Finally,the confidence evaluation is added during the tracking process,when the confidence level does not meet the conditions,the target re-detection module is started,and the target position is re-determined.Three kinds of complex scenes with severe occlusion were selected to test the method.The results show that the proposed method can track the target well and lay a foundation for accurate and continuous tracking of the target.(4)To solve the problem of low processing efficiency due to insufficient utilization of existing computing resources,a target-aware platform based on private cloud is designed and implemented.The system is based on cloud computing.High-speed connected virtual machine clusters are set up through interworking between multiple physical hosts.A cloud-based detection target information awareness system was implemented.The full use of computing resources has been completed,providing an environment for complex large-scale computing such as high-speed,real-time target-aware technologies,realizing the detection,recognition and tracking of different forms of targets.