Design of Real Time Target Tracking Method for Film and Television Video Based on Deep Learning under Visual Communication

Abstract

Video target tracking has gained a lot of interest and applications due to the quick development of computer vision and artificial intelligence. Adaptive modified target tracking approach based on target prediction algorithm and deep reinforcement learning is researched to realize exact positioning of the occluded target and to increase the efficiency, precision, and accuracy of real-time tracking of video targets. And combined with secondary correlation, a multitarget tracking algorithm is proposed to realize target tracking accuracy. The validation experiments are conducted in this research, and the findings indicate that the target tracking effect is at its greatest when the weight adjustment coefficient (p = 0.061) is attained, along with the peak area ratio and similarity of the correlation filtering response reaching their ideal advantage. The target frame only needs to move less than 5 movements in most of the images to successfully capture the target. It is found that the tracking accuracy of the proposed research method has comparable tracking accuracy with the MDNet with optimal performance, while the processing efficiency is improved by 80%, which is an accurate and efficient target tracking method. It is useful as a reference for target recognition in video and has some relevance for target localisation research in subsequent tracking systems.