Research And System Implementation Of Vehicle Behavior Video Recognition Based On Deep Learning

With the rapid development of the road transportation industry in recent years,the frequent occurrence of various traffic violations has put tremendous pressure on traffic supervision agencies.Deep learning-based vehicle behavior video system can effectively and timely monitor vehicle driving safety on various road sections through information and intelligent means.The system can feedback and upload in time for different violations to achieve the purpose of improving supervision efficiency and reducing labor costs.This paper studies the core functions and key technologies of vehicle identification,vehicle violation,and other vehicle behavior determination in the intelligent transportation system and completes the system implementation.The deep learning-based vehicle behavior video recognition system studied in this paper is based on the dual-channel data of the real-time image and local video image returned by the surveillance dome camera,and supports the real-time image capture and forensics function.The first is the pre-processing module,which mainly filters and extracts video images with higher resolution according to the color space and image morphology processing principles.The second is the core functional modules of the system,including target detection and tracking module,violation determination module and predictive snap module.Fast and high accuracy is the key to system implementation for target detection.In this paper,deep learning based on YOLOv3 basic model is used to realize vehicle identification and extraction.Compared with the traditional morphological frame difference method to obtain foreground targets,deep learning has greatly improved the accuracy rate,and can effectively avoid the impact of shadow changes caused by light source changes on vehicle target detection in the monitoring scene.This article adopts the method of maximum overlap area of objects between frames,and uses thresholds to adapt to target vehicle trajectory tracking in different scenes.It can effectively obtain and record the position coordinate information of each frame and each target.This paper presents the recognition algorithm for different violations.Least squares is used to fit curves such as road boundaries,forbidden areas,and various road lines.The obtained boundary information is stored locally in real time for subsequent direct reading,which effectively solves the effect of human error in selecting the boundary reference point.Combining the road boundary data obtained from the fitting and the analysis of the motion trajectory of the target vehicle,it is possible to define the criteria for judging vehicle behaviors such as driving across the yellow line and illegal parking in the no-parking area.Using the pixel coordinate system motion trajectory prediction method to predict the offending vehicle,based on the historical motion trajectory of the offending vehicle,a fractional prediction model is used to predict the best shooting evidence location.This method is not affected by the actual road conditions and the relevant parameters of the surveillance cameras used,it is simple to calculate and robust,and it has good applicability.In the end of this paper,an intelligent tracking and capturing system for traffic vehicle violations is implemented.The hardware equipment part of the system includes dome camera,computer and server;The software part is based on Visual Studio development platform,including dome interface,dome control,image processing,intelligent research and judgment,violation judgment,evidence collection and uploading module.The software part realizes that the system automatically discriminates vehicle behaviors such as violation of regulations and automatically controls the dome camera to capture video.The system conducts tests on frequently-occurring road sections,and captures illegal evidence of illegal activities such as driving across the yellow line and parking in prohibited areas.The test results are good and can effectively identify vehicles and monitor their specific violations.