High Stability Tracking with Sparse Location Information

Abstract

Modern trackers based on traditional image processing or machine learning have made great progress in recent years but still face the problem of error tracking. To make use of these trackers with less error tracking, we propose to combine the object detection and the object localization technology together in our tracking system. The object localization technology can get the positions of moving objects in real time but suffers from low precision, which can compensate this with the modern object detection of high accuracy. The object detection based on deep learning, such as Faster RCNN, which achieves excellent object detection accuracy on PASCAL VOC 2007, 2012 datasets with 300 proposals per image, can get 43 mAP on COCO detection dataset. Object localization devices such as a UWB module will provide us real-time locations of multi-objects with the precision of about 30 cm. These locations can be transformed into pixel coordinates. OpenCV also play an important role in our experiment for providing useful API for camera calibration, coordinate transformation, tracking and so on. With the support of the tools mentioned above, we can develop the state-of-the-art trackers. The major part of the tracker is Long-term Correlation Tracking (LCT). Besides, we provide position information and detection results of our interested objects and try to match them to get reliable positions. By matching the target location from the image tracker and the UWB device, error tracking will be corrected when occurred. The proposed system gives more stable results than existing trackers such as Kernelized Correlation Filters (KCF). This will be helpful for the application scenarios require stable and accurate tracking.