Faster than the fastest : using calibrated cameras to improve the fastest pedestrian detector in the west

Abstract

In this thesis, we translated a state-of-the-art object detector (the Dóllar method) to the C++ programming language and explored ways to use camera calibration to improve its performance by reducing the amount of calculations necessary and to improve the results by taking away false positives. We developed these techniques in the context of pedestrian detection. On data sets more aligned with video surveillance applications (the camera is high in relation to the ground and far from the area where objects are expected to be), we had great results across the board: the amount of scales in the feature pyramid is reduced by about half, the amount of times the classifier is applied is greatly reduced together with the number of false detections, all while keeping the loss in detection coverage manageable. We also tested our detector in one data set that closely resembles the use of detection in robotics or self-driving systems for automobiles (camera closer to the ground plane and parallel to it). The results suggest the method needs adjustments to be applied to this type of setting. Although there was no loss in detection quality and both the number of scales in the feature pyramid and the number of false positives were reduced, the amount of classifier applications seems excessive. To avoid this problem, we need to adjust the Dense Detection phase of our method (subsection 3.2.2) to account for the fact that images created by these camera settings have a bigger range of possible pedestrian heights and more portions of the image are plausible to provide detections.