Improved FairMOT multi-target tracking method for complex scenes

Abstract

At present, the popularization of augmented reality technology on personal terminals is no longer hindered by inherent computing devices due to the substantial improvement in the performance of intelligent terminal processors. However, in the practical application of augmented reality technology, there are still various external disturbances in the real environment, such as perspective changes, motion poses and other potential variables, so it will cause unstable target detection, target registration distortion, virtual and real space superposition of varying degrees of severity. Using TBD strategy as the basis of multi-target tracking technology can further improve the efficiency of multi-target tracking method if it is effectively combined with neural network algorithm. Based on the improved FairMOT neural network multi-target tracking algorithm, this paper mainly studies the specific application of deep learning algorithms in the field of multi-target tracking. Experiments have shown that the improved multi-scale feature fusion network and the optimized loss function of the FairMOT tracking algorithm have good real-time performance and high accuracy for multi-category target detection. The method of stable tracking of multi-category targets is completed in the scene of this paper, which improves the tracking performance, and finally achieves stable tracking through neural network learning and training.