Abstract
In recent years, advancements in computer vision have yielded new prospects for intelligent transportation applications, specifically in the realm of automated traffic flow data collection. Within this emerging trend, the ability to swiftly and accurately detect vehicles and extract traffic flow parameters from videos captured during snowfall conditions has become imperative for numerous future applications. This paper proposes a new analytical framework designed to extract traffic flow parameters from traffic flow videos recorded under snowfall conditions. The framework encompasses four distinct stages aimed at addressing the challenges posed by image degradation and the diminished accuracy of traffic flow parameter recognition caused by snowfall. The initial two stages propose a deep learning network for removing snow particles and snow streaks, resulting in an 8.6% enhancement in vehicle recognition accuracy after snow removal, specifically under moderate snow conditions. Additionally, the operation speed is significantly enhanced. Subsequently, the latter two stages encompass yolov5-based vehicle recognition and the employment of the virtual coil method for traffic flow parameter estimation. Following rigorous testing, the accuracy of traffic flow parameter estimation reaches 97.2% under moderate snow conditions.