An Efficient Multi-Object Tracking Framework for Embedded Vision via Multi-Feature Fusion

Abstract

Despite significant progress in single-object detec tion, existing methods still face substantial technical challenges when deployed in complex and dynamic environments. Moreover, most models lack the capability to maintain stable identity information over time, leading to limited applicability in sce narios requiring persistent tracking and reliable spatiotemporal association. These limitations become even more pronounced on resource-constrained embedded platforms, where computational efficiency and real-time performance must be carefully balanced against detection precision and robustness. To address these chal lenges, our research presents an efficient multi-object tracking framework that combines SSD-based object detection with a multi-feature fusion association strategy. The proposed method leverages complementary appearance, motion, and geometric cues to construct a unified similarity matrix, enabling robust inter-frame correspondence and reducing identity switches under challenging conditions. A Kalman-based motion model and a complete track-management scheme further enhance robustness against short-term occlusions and intermittent detection failures. The system is implemented on a resource-constrained Raspberry Pi 3B platform, with dedicated optimizations to ensure real-time performance. Extensive experiments on the VOT2017 benchmark demonstrate that the proposed approach achieves a favorable balance between accuracy, stability, and computational efficiency. These results highlight the practicality of deploying the frame work in real-world embedded vision applications and provide a strong foundation for future extensions incorporating learned temporal models and lightweight re-identification modules.

Keywords