Vehicle Speed Estimation Based On 3D ConvNets And Non-local Blocks

With the rapid development of the social economy,the car has become a neces-sary means of transportation for daily travel,thus the traffic flow is increased.Traffic congestion is common during holidays.In order to alleviate this phenomenon,the gov-ernments need to build a set of the intelligent traffic monitoring system,and the speed estimation is one of the significant problems.Many existing approaches to this problem are based on camera calibration.How-ever,camera calibration-based methods have many shortcomings.First,camera calibration-based methods are sensitive to the environment,which means the accuracy of the results compromise in some situations where the environmental condition is not satisfied.Be-sides,most camera calibration methods require manual measurement of some physical data and have limitations in camera placement.Furthermore,the accuracy of the speed estimate heavily relies on the accuracy of algorithms for detection and tracking.As an attempt to ameliorate the aforementioned limitations of camera calibration-based methods,we propose an alternate end-to-end method based on 3-dimensional convolutional networks(3D ConvNets).The proposed method bases average vehicle speed estimation on information from video footage.Our methods are characterized by the following three features.First,we use non-local blocks in our model to better capture spatial-temporal long-range dependency.Second,we construct a multi-scale convolutional network.This network extracts information on various characteristics of vehicles in motion.Third,we use optical flow as an input in the model.Optical flow includes the information on the speed and direction of pixel motion in an image.This network extracts information on various characteristics of vehicles in motion.The pro-posed method showcases promising experimental results on a commonly used dataset with mean absolute error(MAE)as 2.71km/h and mean square error(MSE)as 14.62.In a word,the speed estimation method based on 3D ConvNets proposed in this paper is an end-to-end method and applicable to various camera views without manual intervention.