Comparison of two human detection algorithms that apply Bayesian and Neyman-Pearson test criteria using infrared images

This thesis focuses on the application of two human detection algorithms that apply Bayesian and Neyman Pearson test criteria technique for human detection on Long Wave Infra Red (LWIR) images in non-urban environment. These two criteria use threshold values in order to separate regions of interest into human and nonhuman. The selected regions of interest from detection are classified using the Support Vector Machine (SVM).;The SVM with different kernel functions (Linear, Quadratic, and Polynomial) will be used to classify the two human detection algorithms. The performance of the classifiers will then be evaluated separately using a performance evaluation curve, Receiver Operating Characteristic (ROC), to find out which kernel is the best performer. The results from the two detection algorithms will then be compared based on their ability to detect humans using LWIR images in non-urban environment.;My proposed contribution applies the Neyman-Pearson criterion test in designing a human detection algorithm that would maximize the probability of detection with constrained probability of false alarm. The humans are detected within a test image by the use of a 50x20 rectangular template containing image intensity data with selected mean and standard deviation as threshold values. The mean and standard deviation threshold values are selected from the ROC curve after the application of the likelihood ratio test with known prior information on training data to classify the data. This rectangular template is used to classify a region as human or non-human by sliding it over a test image. Compared to the detection algorithm by [3] that applies the Bayesian criterion and assume equal prior for human and nonhuman classes and gaussian distribution to derive a threshold expression, the application of the Neyman-Pearson Test Criterion gave a better detection results using our test images from the Jet Propulsion Laboratory.;A total of 160 LWIR images from Jet Propulsion Laboratory (JPL) were used in this thesis work.