Video Object Detection with MeanShift Tracking

Video object detection, a basic task in the computer vision, is rapidly evolving and widely used in various real-world applications. Recently, with the success of deep learning, deep video object detection has become an important research direction. Although existing deep video object detection methods have achieved excellent results compared with those of traditional methods, they ignore the motion laws of objects and are hard to improve the detection performance of the fast moving objects suffering from deteriorated problems such as the motion blur, video defocus, object occlusion and rare poses. To address this limitation, we add the object trajectory information into the process of the video object detection and devise a novel deep video object detection method which utilizes the MeanShift algorithm to guide the deep neural networks to enhance the video object detection performance. The experiments on ImageNet VID dataset validate that the proposed method can improve the recognition performance of fast moving objects with taking into account the motion laws of objects.