Evaluation on Integrated Multi-Target Tracking and Trajectory Prediction Based on Multiscale Feature Fusion

The importance of visual systems in daily life is self-evident, such as ubiquitous monitoring systems, tracking and shooting of some sports events, and even radar, sonar, and other systems for civilian and military use. The use of computer vision to replace human visual systems is a trend in various fields today, thanks to the development of computer technology and artificial intelligence technology. Among them, multi-target tracking and trajectory prediction are important directions that visual systems have always been unable to avoid. Therefore, this article would study them through multi-scale special fusion methods, attempting to optimize them. The general idea of this article is to quantitatively compare the reconstruction results of experimental images, conduct experiments on this basis, and use multi-scale feature fusion modules to participate in the experiment. The main tool of this article's experiment is the MFSA module, and finally detect and fill in some uncertainties that may exist in trajectory prediction, using trajectory prediction equations. After a series of screening and comparison experiments, this article ultimately found that MFSA, as one of the methods for multi-scale feature fusion, can improve the MOTA index by 79.91%, while the ordinary visual reconstruction system, which can only improve by 73.46%, is much simpler compared to it. In addition, MFSA can accurately locate multi-scale features in the image and provide visual feedback to the user.