Research On The Van Recognition And The Door Monitor Based On Deep Learning

With the rapid development of logistics industry,logistics safety transportation is very important to logistics companies and society.In the process of logistic transportation,if the door of the van is not closed properly,it will cause unnecessary loss and increase the probability of traffic accidents.Therefore,it is very necessary to study the detection and identification of the important carrier van in the logistics transportation industry.At the same time,it is also convenient to monitor transportation and provide guarantee safety transportation.This paper mainly realizes the detection of the door state of the van,namely the calculation of the door opening and the realization of the identification of the van is the basis of the calculation of the opening.The target detection and recognition technology based on deep learning has incomparable advantages compared with the traditional recognition algorithm.Select the yolo-v3-tiny network as the network model for van recognition.A total of 3018 images were annotated as the training data set,and the training parameters were set by referring to the network model of yolo-v3 to train model 1.Improve the structure of yolo-v3-tiny network,add 3*3,1*1convolution layer in the feature extraction network,configure the same training parameters asyolo-v3-tiny,and train model 2.Through the loss curve analysis of the two models,it is concluded that when the training time increases by 3 hours,the minimum loss value of the network model with increased network structure is less than that of the unimproved network model.The two models were tested with the same test data set,and the correct detection and recognition rate of van was improved by 3.58%.The improved network model can better distinguish the target van model from other models,such as minibus,car,bus,truck and so on.In this paper,the use of the scene for the yard exit,on the basis of vanidentification,van door state detection.As the shooting angle is fixed,four pairs of pixel coordinates are used to determine the perspective transformation matrix,so as to realize the perspective transformation of the picture.Image preprocessing,such as image grayscale,image binarization,image expansion processing and image edge detection,is carried out on the converted images to achieve the extraction of the contour of the car door.Hough transform is used to detect the straight line of the preprocessed image.By setting the threshold of the straight line slope and the straight line length,the contour line of the compartment door is screened out.Then through the two line slope to determine the angle between the two lines of mathematical calculation method to achieve the compartment door opening calculation.The processing calculation of several different compartment door opening area,compartment door opening calculation value and the real value error of compartment door opening is less than or equal to±1°,the realization of compartment door opening calculation.