We integrated a sensor hardware test-bed using scientific grade, commercial off-the-shelf (COTS) technology and developed supporting software to enable rapid prototyping. The validity of this test-bed and associated software was demonstrated through the delivery of a ground-based multispectral visual surveillance prototype for improvised explosive devices (IED) detection using electro-optical (EO) and short-wave infrared (SWIR) cameras. Software developed to support the test-bed included modules for image acquisition, preconditioning, segmentation, feature extraction, data regularization and pattern recognition. To provide spatially co-aligned data, we optimized a mutual information-based image registration algorithm to improve its convergence rate and benchmarked it against the established simplex method. For four different multimodal test image sets, our algorithm convergence success improved by 15 to 40% as compared to the downhill simplex method, albeit with an approximately four times higher computational cost. Additional strategies, such as bit-depth reduction, image down-sampling and gradient-based regions of interest (ROI) selection, were systematically evaluated and led to the registration of high resolution images at nearly 60 times faster than the standard approach.
To automatically identify IED in the acquired multi-spectral imagery, four different pattern classifiers were tested; Bayes, k-nearest neighbor (knn-NN), support vector machines (SVM) and our novel piece-wise linear convex-hull classifier. Initial tests with the convex-hull classifier using simulated data indicated significant reduction in error rates of up to 89% (p=3e-6) when compared to the Bayes classifier. Subsequently, each of the four classifiers was tested using the IED data set that consisted of 154 different intensity-based and content-based features extracted from the EO and SWIR imagery. Salient features were selected using receiver operating characteristic (ROC) curve analysis and a wrapper-based process that used minimum error rate as the criterion function. A stratified 10-fold cross-validation analysis was used to compare classifier performances using t-tests with a 95% confidence interval. Under different operating conditions (OC), the generalized classification error rate ranged from 13.0% (±1.2) to 29.6% (±2.7). For six of the eight acquired data sets, the convex-hull classifier provided equivalent or lower (as compared to at least one other classifier) generalized error rate, used a lower number of features (at most 3), reduced training time (as compared to the knn-NN and SVM classifiers) and relied on lower function complexity. Concurrently, the class-label assignment time using the convex-hull approach was a factor of 10 lower than the Bayes classifier and a factor of 100 lower than the knn-NN and SVM classifiers. Considering the attributes of a good classifier, the convex-hull approach provided the overall best balance between the different measures of performance.