Research on CNN and Phase Correlation Fusion Algorithm for Vehicle Stopping State Detection in VINS System with Dynamic Scenarios

Abstract

Vehicle stopping state detection in dynamic scenes is a key technology for autonomous driving and Intelligent Transportation Systems (ITSs), playing a crucial role in enhancing vehicle localization accuracy and safety. This paper proposes a vehicle stopping state detection method based on a fusion algorithm combining Convolutional Neural Networks (CNN) and phase correlation, aiming to address the challenges of detection accuracy degradation caused by moving object interference and complex environmental changes in dynamic scenes. The method extracts static background features and filters out dynamic objects using a lightweight CNN model, while leveraging the Phase Correlation Algorithm (PCA) to compute global Motion Vectors (MVs) between consecutive frames, thereby achieving high-precision detection of vehicle stopping states. The experimental outcomes reveal that the suggested technique shows outstanding efficacy on both public datasets and real-world scenarios, achieving a dynamic object detection accuracy of over 75% with the computational load greatly reduced to 2.3 Floating Point Operations (FLOPs). The entire fusion algorithm can improve the accuracy of vehicle stopping state detection to 95.6%, while meeting real-time requirements (inference time of 22 ms per frame). Additionally, the method maintains strong robustness under low-light and adverse weather conditions, providing a new solution for optimizing Visual-Inertial Navigation Systems (VINS) in dynamic environments. The research outcomes of this paper not only provide theoretical support for stopping state detection in autonomous driving systems and lay a technical foundation for the realization of ITSs.