A Novel Video Quality Evaluation Method Based on Coarse-to-fine Strategy in Traffic Scenario

Video Quality Evaluation (VQE) is one of the key technologies for traffic monitoring system since the massive traffic monitoring may produce various distortions in the process of video recording, compression and transmission, which may cause the degradation of the monitoring picture quality. In literature, most of the traditional VQE methods are based on the pixel level, which indicates a limited number of video quality issues. However, the complexity generated by the constant changes of video images in actual scenes can easily lead to misjudgment and omission of this technology. In view of this, current deep learning methods propose to use object detection to solve it. However, since the features of deep learning mainly behave as problems, the performance is limited when there are non-quality targets that approximate the learned texture. Due to the openness of video quality issues, that is, there are various quality situations, the lack of datasets is also the focus of current research. To this end, this study constructs a dataset that can cover a variety of common video quality issues based on traffic monitoring, and proposes a coarse-to-fine quality evaluation method around this dataset. The method is mainly composed of a quality detection module and a quality judgment module. Experimental results show good performance on our proposed dataset.