Non-maxima supression

Abstract

Non-Maxima Suppression is a very important part on the object detection process. When searching for objects in and image several points are usually found as objects but some of them are not really objects, Non-Maxima Suppression (NMS) consists in select which of those maximas are really objects and suppress those that are not. In this thesis different Non-Maxima Suppression for Hough based images methods have been tested, the methods are Gall, Wenzel, Thresh and Islands. Those methods work with Hough images, which are greyscale voting images where white is high probability to be a maxima and black the opposite. The methods have different characteristics and different ways to act depending on the dataset and the used threshold. To compare those methods TUD Pedestrians and TUD Campus datasets were used to obtain the Precision and Recall charts. The different methods had a very different response to the chosen threshold or to the different datasets which in these cases consist in people walking in several places. Gall method is the most robust one, Wenzel and Thresh have almost the same response to the datasets and Islands with a different way to work has a different response than the other ones. The conclusion was that the correct selection of the threshold and the overlap for the different methods and datasets is very important to achieve good results at object detection, what makes creating algorithms for different environments and illuminations much more difficult than for known environments and controlled situations.