Binarized normed gradients for object detection

Object tracking and detection have been one of the most important and active research areas in the computer vision field. A large number of tracking and detecting algorithms have been proposed in recent years, and those algorithms have solved problems in object occlusion, fast motion, deformation, scale variation, or illumination variation. However, there are still some serious problems in heavy occlusion. In this paper, base on TLD (Tracking-Learning-detection) framework, we use binarized normed gradients (BING) to search objects by objectness scores which is operated through a linear SVM(Support Vector Machine) model. Firstly, we resize the input window to different quantized sizes (e.g. 8 × 8) and calculate the normed gradients of each resized image. Then according to the different gradient model of the object and background in the fixed window, we can quickly and accurately locate the target object. Finally, the binarized normed gradients (BING) is used for efficient objectness estimation. The experiment results show that our method can solve object heavy occlusion and improve object detection rate.