Target detection in FLIR imagery using independent component analysis

Target detection in forward looking infrared (FLIR) imagery is a challenging problem. Existing target detection algorithms are computation intensive and involve an inefficient preprocessing step [1-4]. In this thesis, a new algorithm is proposed for target detection in FLIR imagery using independent component analysis (ICA). For training purposes, several real image frames from a FLIR sequence are employed. The dimension of the training regions (windows) is chosen depending on the spatial size of the target of interest in the FLIR sequence. After applying the ICA transformation to these training images, a weight matrix and an ICA matrix are obtained. The two matrices thus obtained are then employed to obtain a transformed matrix of the input image with enhanced features. In the proposed algorithm, a moving window with a predefined size is constructed in the first frame such that at every pixel a vector is generated for the corresponding window. Then the value of the cosine is calculated between that vector and the target vector of the training matrix. Thereafter, the highest value is considered to identify a potential target region in the scene. A detailed simulation software is developed using the MATLAB software package to investigate the performance of the proposed algorithm. Test results using real life FLIR imagery show promising results for practical applications.