Due to factors such as manual operation, strong winds, and object obstruction, the field of view in traffic monitoring scenes often undergoes dynamic changes. These changes not only hinder the monitoring system’s ability to accurately perceive the traffic environment but also undermine its reliability. Existing methods typically focus on detecting whether a scene has changed, without addressing the specific degree or type of these changes, limiting their practical applicability. To overcome these limitations, we propose an automatic detection method for traffic scene variations with stable scene representation features and quantifiable degrees of change. Specifically, our framework combines semantic understanding of traffic scenes with spatiotemporal change analysis. A deep semantic understanding model is developed to extract features representing the stable state of traffic scenes, while a novel quantification method measures the degree of semantic feature changes. Different types of changes are classified using a threshold-based approach. Our method can automatically detect significant changes in traffic monitoring scenes under various environmental conditions, including camera rotations and lens zooming. Experiments conducted in real-world settings demonstrate the robustness and effectiveness of the proposed approach. The deep semantic understanding model achieves an average detection rate of 95.3%; camera perspective change detection reaches 97.04%, surpassing traditional scale-invariant feature transform-based methods by 12.33% and convolutional neural network–long short-term memory-based methods by 11.46%.
Lu et al. (Thu,) studied this question.