Feature optimization and creation of a real time pattern matching system

State of the art algorithms for people or vehicle detection should not only be accurate in terms of detection performance and low false alarm rate, but also fast enough for real time applications. Accurate algorithms are usually very complex and tend to have a lot of calculated features to be used or parameters available for adjustments. So one big goal is to decrease the amount of necessary features used for object detection while increasing the speed of the algorithm and overall performance by finding an optimum set of classifier variables. In this paper we describe algorithms for feature selection, parameter optimisation and pattern matching especially for the task of pedestrian detection based on Histograms of Oriented Gradients and Support Vector Machine classifiers. Shape features were derived with the Histogram of Oriented Gradients algorithm which resulted in a feature vector of 6318 elements. To decrease computation time to an acceptable limit for real-time detection we reduced the full feature vector to sizes of 1000, 500, 300, 200, and 160 elements with a genetic feature selection method. With the remaining features a Support Vector Machine classifier was build and its classification parameters further optimized to result in less support vectors for further improvements in processing speed. This paper compares the classification performance, of the different SVM's on real videos (some sample images), visualizes the chosen features (which histogram bins on which location in the image search feature) and analyses the performance of the final system with respect to execution time and frame rate.