Improving object position estimation based on non-linear mapping using Relevance Vector Machine

Abstract

The objective of the proposed work is object position estimation, in which the system, after training with examples of images including objects such as cars, should be capable of indicating accurately by coordinates. The method is different from simple object detection, since it uses the context, i.e. the whole image. The key idea is to take an approach with Relevance Vector Machine (RVM) since it leads to sparse models and theoretically better performance is expected compared to previous proposals. The RVM mapping was done first as a training stage, in this case by using the same image database as the conventional method used as comparison with a previous Support Vector Regression proposal, where cars in different positions and sizes are included, and with exact coordinates given explicitly to the system, after this, it can perform without previous training.