A Camera Movement Guidance Method based on Multi-Object Tracking

Abstract

Multi-object tracking (MOT) attracts great attention in computer vision while playing an increasingly essential role in manufacturing and life. There is currently an urgent requirement for practical implementation of MOT algorithms to guide production in industry. Nevertheless, the algorithm guiding camera movement is prone to be naive in reality to some degree, which is not conducive to handle complex situations and may cause damage. In this paper, we propose a deep-learning-based MOT method to guide camera movement. In particular, jointly learnt detector and embedding model (JDE) is adopted to extract the features of live stream through data preprocessing and training, which detects the spacial and temporal information of objects, for instance, pedestrians. Moreover, we propose a cognitive-based network segmentation method which makes edge-cloud collaboration possible. Additionally, object location information provided by deep learning allows object clustering and weight allocation, followed by utilizing PID algorithm to guide camera motion. Our method is compared with conventional model through several metrics, especially Euclidean Average Distance, which indicates the effectiveness, reliability and robustness of our model.