| In order to investigate the cause of a traffic accident,sometimes we need to make a judicial appraisal to the accident during the investigation stage.The traffic police then divides the responsibility of the accident according to the appraisal issued by the judicial appraisal institution.At present,the appraisal procedure for the vehicle recognition task is still relying heavily on the method of manual comparison – by studying the target vehicle in a video captured from non-fixed angles of view.Unfortunately,there are still numerous disadvantages associated with the current method that would potentially hinder the matching process.For example,mismatching in shots taken from different shooting angles,low image resolution and parts of the vehicle body being covered will all possibly add difficulty to the checking process.Therefore,the existing recognition method is not suitable for the judicial appraisal of road traffic accidents.In contrast to the main vehicle recognition method implemented nowadays,extracting vehicle characteristics from a single vehicle picture through the method of deep learning and then classifying the characteristics into vehicle types accordingly seem to be a more sophisticated solution.For the purpose of applying judicial appraisal on traffic accidents,this thesis proposes to use images showing vehicle in both its entirety and its parts as the data basis for vehicle recognition.The work of this thesis is as follows:1.This thesis establishes two new image databases as a response to issues such as partial vehicle being covered in the traditional database and also to differentiate from the current database that only includes images of whole vehicles.The firstdatabase,which comprises of 35 vehicle makes and 135 models gathered from online research,includes nearly 10,000 raw images detailing both whole vehicles and vehicle parts.Image processing methods,such as mirroring image,noise and blur editing,are then adopted to produce final database of approximately 80,000 images.The other database includes nearly 1,600 images of whole or partial Volkswagen models derived from surveillance videos used in prior judicial appraisals.2.To more accurately detect vehicles in traffic accident judicial appraisal,the overall framework of obtaining relevant parameter information of the target vehicle is designed as follows.First,classify its vehicle make.Next,classify its body type(coupe,hatchback,sedan,sports,SUV,MPV,or etc.)and models.Finally,using the three classifiers of the target vehicle obtained above,we match check and retrieve the corresponding entries of its parameter item information in the self-built text database.3.Using the four classical convolutional neural networks(AlexNet,Vgg16,GoogleNet v3,and ResNet50)as deep learning models,the transfer learning method is applied on the two databases mentioned above to detect vehicle makes.The recognition accuracy rate achieved on the 35 vehicle makes in the online picture databases came up to 88.94% and 99.49% for 10,000 and 80,000 records respectively.In addition,recognition accuracy rate achieved on the Volkswagen images from the surveillance videos database also revealed to be as high as 90.16%.4.For further improvement of detection accuracy,cropping of the vehicle from the whole image is necessary.This is achieved based on the following.First,use ImageLabeler to label the vehicle location in 897 pictures from Compcars image database.Next,cluster the ratios of vehicle width and height in pictures using K-means cluster.Then,apply the clusters to Faster R-CNN and finally use the modified Faster R-CNN results to detect and crop out the 58 different Volkswagen models in Compcars database to remove background noise in the photos.Once cropped,the transfer learning method is adopted on the updated photos for better detection.The recognition accuracy rate for classification of body types(coupe,hatchback,sedan,sports,SUV,MPV,or etc.)and models under the revised method improved from 83.56% to 93.01% and 79.74% to 89.70% respectively. |
PDF Full Text Request