Multi-Object Tracking Algorithm Based On Feature Association And Feature Enhancement

Multi-Object Tracking is one of the branches of computer vision,which has a broad application prospect in intelligent security,human-computer interaction,intelligent robots,and automatic driving.Multi-Object Tracking algorithm is generally based on object detection,and detection based Multi-Object Tracking algorithms can be divided into two categories: One-Shot Multi-Object Tracking algorithm and Two-Steps Multi-Object Tracking algorithm.One-Shot Multi-Object Tracking algorithm is widely studied and applied because of its fast speed and high accuracy.Detection and extraction of appearance features are two sub-tasks of One-Shot Multi-Object Tracking algorithm.The tracking performance of the algorithm are limited due to the conflict between the two sub-tasks in model optimization.Furthermore,the MOT algorithm suffers from the issue of identity switching too frequently,which has an impact on the tracking effect.As a result,this article conducts research on the above issues,taking the sub-tasks conflict of One-Shot Multi-Object Tracking as the starting point,and studies how to relive the conflict and extract high-quality appearance features,and constructs a high-performance Multi-Object Tracking algorithm.The main work of this article is as follows:(1)Construct a One-Shot Multi-Object Tracking framework.The performance of Multi-Object Tracking algorithm is improved by adding Re-Identification branch after anchor free object detection network and using data association method of hierarchical matching of multi-quality detection results.(2)The feature association module is introduced into the One-Shot Multi-Object Tracking algorithm.This article analyzes the conflict between two tasks of detection and extraction of appearance feature,and designs feature association module.The experiments show that the feature association module can effectively alleviate the conflict in One-Shot Multi-Object Tracking and improve the tracking performance.(3)Feature enhancement module is designed in appearance feature extraction task.This study examines the issue of identity switching,which is common in MOT algorithms.The main cause of the issue is a lack of expressiveness in facial features.The representativeness of appearance features can be improved by extracting space and channel attention with large kernel convolution.The feature enhancement module can significantly reduce identity switching in MOT scenarios and improve the coherence of target trajectories in images,according to experimental results.Ablation experiments are carried out for the proposed feature association and feature enhancement of the algorithm.The experiments prove that both feature association and feature enhancement can bring a gain effect to the tracking performance,and the tracking effect on several datasets exceeds the current SOTA Multi-Object Tracking algorithm,which has high practical utility.