A Robust Adaptive Temporal Attention Mixture Network for Multi-Object Tracking and Segmentation

Abstract

Traditional video object tracking and segmentation methods including Gaussian Mixture Models (GMM) and fuzzy morphological filtering produce unstable results when operating under conditions of changing illumination, camera movement and object blocking. The current deep learning and attention-based methods have enhanced temporal stability but most existing pipelines continue to operate in separate stages for segmentation and denoising and tracking which results in progressive error accumulation and decreased real-world performance consistency. The research introduces ATAM-Net (Adaptive Temporal Attention Mixture Network) as an end-to-end framework which unifies adaptive Gaussian mixture modeling with temporal attention-based denoising and differentiable object association into one system. The ATAM-Net system learns particular illumination parameters for each pixel through its adaptation process and temporal attention minimizes frame-to-frame noise and appearance-aware embeddings maintain identity during fast motion and occlusion. The SportsMOT dataset serves as the exclusive testing ground for ATAM-Net to prove its ability to produce stable visual segmentation and smooth temporal coherence and precise multi-player tracking in active sports environments. The qualitative visual results show clear object edges and minimal flicker. These results indicate that ATAM-Net provides high-motion environments such as sports analytics, reliable and interpretable approach for multi-object tracking and segmentation in complex and real-time video understanding.