| It is a practicable approach to use the infrared radiation of the space time-sensitive moving targets,due to the targets’ extremely high temperature compared to space background.The infrared-radiation-signal-based analysis of space time-sensitive moving targets mainly contains two parts which are the detection of key points,i.e.,obtaining the time when fatal events happened,and the classification of the targets themselves.The detection of key points is helpful for determining the targets’ status and can provide extra information for classification.On the other hand,the classification of targets is the prime objective,which enables staffs to take rational actions as soon as possible.Our research therefore will focus on the detection of key points and the classification of targets and be mainly composed by the following four parts.(1)We will propose a key point detection algorithm based on stacked LSTM obtaining precision of 0.963,recall of 0.951,F1-score of 0.957 and segmentation covering metric of 0.773 on our own dataset,and a new metric to evaluate the key point detection performance with different demands of temporal precision.We will also propose the improvements of labeling,data augmentation,loss function and postprocessing along with our key point detection algorithm.(2)We will propose a target classification algorithm based on the key point local features which are composed by time domain features,time-frequency domain features,and deep features.Our algorithm compresses the key point local features using FRESH algorithm and principal component analysis,and then classifies them with a random forest,which obtains precision of 0.972,recall of 0.967,and F1-score of 0.967.(3)We will propose a multitask neural network for key points detection and targets classification and elaborate the approach of deploying the model on software platform as well.Due to multitask learning and key point attention,our algorithm can detect key points and classify the targets simultaneously with the similar performance to the former single-task algorithms.It achieves precision of 0.953,recall of 0.954,F1-score of 0.953 and segmentation covering metric of 0.721 on key point detection task,and precision of0.982,recall of 0.982,and F1-score of 0.980 on targets classification task. |
