This paper explores the fault tolerance of deep learning-based multi-object tracking under transient hardware faults. Such faults, often caused by radiation-induced bit flips or temporary electrical disturbances, can corrupt model parameters or intermediate activations during inference, leading to potential performance degradation. Unlike prior work focused on image classification, we examine the more complex task of tracking using ByteTrack with YOLOX and YOLO-Nano backbones. Our GPU memory fault injection framework reveals that common mitigation methods like activation clipping are ineffective in this context. We propose a novel Temporal Consistency Filter (TCF) leveraging frame-to-frame similarity to detect and correct faulty feature extractions. TCF significantly improves tracking stability, boosting MOTA by over 9\% and reducing identity switches by 17%, emphasizing the value of temporal consistency in robust tracking systems.

Fault tolerant multi-object tracking via temporal consistency filtering / Milazzo, Rosario; Fosson, Sophie; Morra, Lia; Sterpone, Luca. - ELETTRONICO. - (In corso di stampa). (Intervento presentato al convegno 23rd International Conference on Image Analysis and Processing (ICIAP) 2025 tenutosi a Roma, Italia nel 15-19 Settembre 2025).

Fault tolerant multi-object tracking via temporal consistency filtering

Rosario Milazzo;Sophie Fosson;Lia Morra;Luca Sterpone
In corso di stampa

Abstract

This paper explores the fault tolerance of deep learning-based multi-object tracking under transient hardware faults. Such faults, often caused by radiation-induced bit flips or temporary electrical disturbances, can corrupt model parameters or intermediate activations during inference, leading to potential performance degradation. Unlike prior work focused on image classification, we examine the more complex task of tracking using ByteTrack with YOLOX and YOLO-Nano backbones. Our GPU memory fault injection framework reveals that common mitigation methods like activation clipping are ineffective in this context. We propose a novel Temporal Consistency Filter (TCF) leveraging frame-to-frame similarity to detect and correct faulty feature extractions. TCF significantly improves tracking stability, boosting MOTA by over 9\% and reducing identity switches by 17%, emphasizing the value of temporal consistency in robust tracking systems.
In corso di stampa
File in questo prodotto:
File Dimensione Formato  
ICIAP.pdf

accesso riservato

Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: Non Pubblico - Accesso privato/ristretto
Dimensione 419.32 kB
Formato Adobe PDF
419.32 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3002484