Research On Key Technologies For Video Quality Evaluation In Traffic Monitoring Scenarios

Video surveillance is a vital part of the traffic management system,and the quality evaluation of surveillance video images is an important means of maintaining effective monitoring at all times.However,various distortions may occur in the process of massive video recording,compression and transmission,which may cause the degradation of the monitoring picture quality.Therefore,intelligent evaluation technology is urgently needed to provide quality feedback on surveillance videos.Earlier research focused on traditional solutions,such as the parametric evaluation technology that uses image pixels to compare and judge(it is necessary to evaluate indicators such as signal-to-noise ratio with reference to a given image),but the complexity of the technology produce in constant changes of video images in actual scenes can easily lead to misjudgment and omission of the technology.Compared with the parameter-based method,although the parameter-free evaluation technology reduces the requirements for providing reference image,however,it needs sample diversity to support the final model training and mainly focuses on the detection of limited noise types.Therefore,it is inevitable to cause misjudgment of the approximate target.Inspired by the principle of object recognition,this study equates the problem of quality determination as the problem of object detection and classification.That is,in view of the quality problems in complex traffic scenes,such as occlusion(leaves,billboards,etc.),electronic interference or signal instability,an integrated video quality evaluation method based on fusion of deep learning features and traditional features is proposed.The main contributions are as follows:(1)Aiming at five common video quality problems,namely mosaic,streak,snowflake,occlusion and blur,a deep learning-based method for evaluating video quality of traffic surveillance without reference is proposed.The model studied directly learns semantic-level features,and adds feature pyramid structure and attention mechanism,and finally enables the model to learn more feature information through feature fusion.In addition,considering the unbalanced characteristics of the samples,the model adopts the Focal Loss function to reduce the impact of unbalanced samples.(2)Using the idea of ensemble learning,the SVM model is obtained by integrating and training texture features such as HOG,so that the model evaluation results can be refined and optimized.Finally,constructed a quality evaluation system with simple interface and easy to use by Qt,based on the above ideas.It is verified on a real traffic surveillance video dataset,and the results show that the research model has good quality evaluation results for five typical quality problems: mosaic,stripe,snowflake,occlusion and blur.Compared with the basic network,the accuracy rate of the improved deep learning model is increased by 3.95%;the fusion model after adding SVM,compared with the deep learning model,the accuracy rate is increased by 2.12%.Compared with the evaluation method using the classical object detection network,the fusion model studied in this paper has higher accuracy and better effectiveness.