IR Small Target Detection And Prediction With ANNs Trained Using ASSET

Abstract

This paper presents and solves a constrained realworld problem of using synthetic data to train artificial neural networks (ANNs) to detect unresolved moving targets in wide field of view (WFOV) infrared (IR) satellite motion imagery. Objectives include demonstrating the use of the Air Force Institute of Technology (AFIT) Sensor and Scene Emulation Tool (ASSET) as an effective tool for generating IR motion imagery representative of real WFOV sensors and describing the ANN architectures, training, and testing results obtained. A 3-dimensional convolutional network (3D ConvNet) and long short term memory (LSTM) network were used within the same pipeline to achieve the goal of IR small target detection and prediction. The 3D ConvNet was generally able to predict target locations with a mean absolute error (MAE) median of 0.31 using the temporal and spatial features of motion imagery. The LSTM predictions were based on previous target locations and incorporating the LSTM predictions to future frames resulted in a MAE median of 0.45. This paper uses a deep learning model with both convolution and LSTM components that performs well on IR small target tracking, which would otherwise be difficult to detect in individual frames.