Machine Learning for Missile Streak Detection and Localization

Abstract

This paper presents a machine learning approach using Neural Networks for the detection and localization of a single streaking target from an optical sensor's Focal Plane Array (FPA). Current state-of-the-art methods utilize model-based probabilistic techniques, such as the maximum likelihood method. The image data for training and the concomitant ground truth for detection and localization is generated via simulation due to lack of sufficient amounts of real-world data. The images were generated assuming that the target's point spread function (PSF) is Gaussian and moves during the FPA's integration time. Also, a noise model for each optical pixel on the FPA is used that is consistent with a Poisson model of the number of non-target originated photons. Our approach divides the problem into two parts: streak detection and localization. To simplify neural network architecture selection and training, we split the large images into smaller sub-images. The noisy sub-images were preprocessed by an autoencoder with Convolutional layers, which preserved the target intensity. Next, we used the denoised sub-images from the autoencoder to detect streaks. For localization, the Neural Network was given the original sub-images to predict the pixel locations of a streak's start and end points. All three processing steps utilized Convolutional layers. The performance of the detection model was evaluated in terms of confusion matrices at different noise intensities, as well as the Receiver Operating Characteristic (ROC) curves at a constant noise intensity. For the localization model, we evaluated the Mean Squared Error (MSE). These results were compared to the Generalized Likelihood Ratio Test (GLRT) for detection and an existing matched filter-based maximum likelihood method for localization. We found that the machine learning models were 340 times faster for detection and approximately 360 times faster for localization on full size 256x256 images. In addition, the detection accuracy of the machine learning model was substantially better than the maximum likelihood method (area under the ROC of 0.94 versus 0.85) and the MSE was significantly lower (0.0261 versus 0.0622 on 32x32 sub-images and 0.0351 versus 0.243 on 256x256 original images). These results are significant for time-critical systems, such as missile defense systems (e.g., C-RAM). Aside from missile defense, the approach presented here can be applied to counter many other weapons that exhibit similar linear streak trajectories, such as torpedoes. In this case, sonar images would be used rather than images from an optical sensor.