(A) robust real-time target tracking framework for complex environment using a stereo Camera

Abstract

Deep learning-based object detection and object tracking algorithms have enabled robots to perform many tasks. However, if the robot tracks the target, it becomes a difficult problem by changing the camera position, the object position, and the surrounding environment. In a complex environment, it is a challenge for robots which has limited computing resources to track an object in real-time. In this research, real-time tracking by detection frameworks using a stereo camera is proposed for single object tracking mobile robots. The proposed framework reduces inference time by implementing the object detector on TensorRT. The color histogram, the object bounding box, and the 3D location calculated from depth map are used for object identification. The movement of the object is compensated using the Kalman filter, and the movement of the camera is compensated using egomotion. In addition, it uses an appearance learning method called memory book to remember the change of the target appearance, the weight decay step adaptively adjusts the weights of features for object association to increase the re-identification. The proposed algorithm was verified using various datasets, and it was confirmed that a single object can be robustly tracked in a complex environment. The proposed algorithm was tested in various dataset environments, and verified that a single object can be robustly tracked in a complex environment.