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.