A Fast and Precise HOG-Adaboost Based Visual Support System Capable to Recognize Pedestrian and Estimate Their Distance

摘要

In this paper,we present a visual support system the visually impaired. Our detection algorithm is based on the well known Histograms of Oriented Gradients (HOG) method, due to its high detection rate and versatility. However, the accuracy of object recognition rate is reduced because of high false detection rate. In order to solve that, multiple parts model and triple phase detection have been implemented. These additional filtering stages were conducted by separate action on different area of the sample, considering deformations and translations. We demonstrated that this approach has raised the accuracy and speed of calculation. Through an evaluation experiment based on a large dataset, we found that false detection has been improved by 18.9% in respect to standard HOG detectors. Experimental tests have also shown the system ability to estimate the distance of the pedestrian by the use of a simple perspective model. The system has been tested on several photographic datasets and have shown excellent performances also in ambiguous cases.