Vehicle Detection Based On Convolution Neural Networks In Infrared Image

The safety issue of vehicle driver operating during night or in awful weather is always catching attention of people.Therefore,in order to improve the safety level of driving,intelligent vehicle safety system has become one of the popular topic of development and research and leading the direction of the whole relevant field.Vision based vehicle detection is the crucial feature of intelligent vehicle safety system,and it is also a significant section in the field of machine vision.When driving in an environment which prone to cause traffic accident,such as dark night,the image based-on visible light will not get enough environmental information.However,comparing to the image based-on visible light,the infrared image can be used as the input of the intelligent system at night or other critical weather situations because it can be used in all-weather situation and no environmental light requirement.The vehicle detection system based on infrared image is a part of the intelligent vehicle safety system,the related research about it is how to avoid traffic accident,this absolutely means a great significance to elimination the potential threat of life and property.The theory and application of machine vision have been greatly developed from the beginning to the present.At the same time,as an important research in the field,target detection has made a huge progress in recent years.Especially with the emergence of the concept of deep learning,the combination of convolutional neural network and target detection task make us see a great progress in the relevant research field.Therefore,the research on the application of convolution neural network in deep learning of vehicle detection in infrared images is the main topic of this thesis.The work in this thesis mainly includes the following aspects.First of all,according to the purpose of this thesis,the infrared images of many vehicles on the road are collected under different time and different scenes,and each picture from the data need to make corresponding annotation files.Some of the images are used as training data to prepare for the model training.The others are used as a test set to test the model after it is trained.Secondly,according to the characteristics of infrared images,the proposed region is extracted on the basis of the selective search algorithm,and the R-CNN model based on Alex Net network is used to detect the vehicle detection in infrared images.After the Image Net data set for pre-training Alex Net network was used in the experiment,and then use the infrared image data set that we maded to fine-tuning training,experimental results show that the network model has achieved 81.5% accuracy,and compared to the traditional detection method,it has increased a lot,and detection time for each picture is 8.6s.Finally,in order to obtain better detection performance,an improved ZF-Net network is proposed.And it is combined with Single Shot Multi Box Detector to achieve the infrared image of the vehicle detection.The same data set is used to train and test,then compared with the experimental results of R-CNN model.The results show that the improved ZF-Net network combined with the SSD model achieves 87.3% of the detection accuracy,the detection speed is more than 0.024 s per picture,close to 42 fps.This result shows that we achieve accurate real-time detection.