Orientation Robust Object Detection Using Histogram of Oriented Gradients

Abstract

We study orientation robust object detection using the HOG feature set. We show that this method provides reasonably well accuracy for detecting objects with varying angle, poses and distance from the viewing plane. We calculate gradient magnitude and orientation of individual cells of the input images and get gradient vector from it. Then we divide the gradients vectors in predetermined bins depending on it's orientation. After that, we normalize the image blocks to get normalized vector. Concatenating all the normalized vectors gives our nal feature vector. Finally we give the feature vector to a SVM to train our detector. Once the detector is trained, it is ready for testing. Our testing results show that the detector can detect objects with recall rate of 83% and precision rate of 97%.