Research On Real-time Identification Method Of Pedestrains And Vehicles Based On R-CNN

Urban road monitoring is an effective means of monitoring road traffic safety,rectifying urban security and preventing various crimes and terrorist violence.In the past,there were many types of road surveillance cameras,and the video data integration work was done manually.It required a large number of professional technicians.Nowadays,with the development of machine vision technology and the advancement of monitoring intelligence,the performance of the algorithm has surpassed humans in many image sensing problems,and has entered the stage of large-scale engineering application,but the comprehensive application is still in its infancy.Therefore,this paper aims to integrate the latest visual technology research results,and build a set of R-CNN(Regions with Convolutional Neural Network)-based video input that can rely on different types of surveillance cameras,which can realize simultaneous detection and identification of human vehicles.In this paper,the existing intelligent monitoring,pedestrian recognition,vehicle license plate detection and other technical fields have been investigated,and the overall scheme is constructed.The whole process is for three modules: the target detection algorithm module,which is responsible for the pedestrian vehicle from the surveillance video.The image information is detected and extracted in real time;the pedestrian recognition algorithm module Re-Identifies the comparison between the pedestrian image and the data inventory image;the vehicle license plate recognition algorithm module performs the license plate recognition on the detected vehicle image.Then,this paper has conducted in-depth research on three algorithms:The target detection algorithm compares the Mask-RCNN and Light-Head RCNN algorithms with other recent achievements,and improves the performance of the algorithm,balancing the accuracy and real-time performance;Pedestrian re-recognition algorithm,based on metric learning method,designed adaptive loss function and combines human analytic algorithm with re-recognition.Based on dynamic self-encoding method,a general self-encoding algorithm is proposed to make the algorithm target the environment.Improved performance in more complex lighting conditions;The vehicle license plate recognition algorithm applies the weakly supervised irregular character detection method,and introduces a hard attention mechanism.At the same time,the license plate sample is generated for model training,and the specific a priori statistical information is used for the classification of the provincial abbreviation recognition rate.The threshold is corrected to improve the accuracy of the license plate recognition of the traveling vehicle.