Car Detection And Vehicle Type Classification Based On Deep Learning

With the rapid development of science and technology, China is in the crucial stage of modernization, at this moment for the majority of the Chinese people, cars are no longer just a kind of transport vehicles, but also an essential part of people’s daily life. Presently, growing traffic congestion, traffic disturbance and traffic accidents have been becoming the stubborn disease which restricts the social improvement and economic development of a region. That how to effectively reduce the incidence of traffic accidents and alleviate the traffic congestion have become an urgent task for social development. Thus, the idea of intelligent transportation was emerged as a result, in which the car detection and vehicle classification are two important parts. Since 2006 deep learning has found a wide application in image recognition, speech recognition and other fields, the results of face recognition are even better than human. Deep learning does not require manual design features and can precisely fit nonlinear input, and these features make it a promising tool in car detection and vehicle classification. This paper is supported by the national natural science foundation of china and the department of transportation project, which is focus on car detection and vehicle type classification. The main work is as follows:1. With aspect ratio of the detection window arbitrary scaling and modifying the network loss function, the modified convolutional neural network model is proposed, which can fit for the nonlinear function of vehicle location regression problems and locate the region of the vehicle in the natural scene images of any size at any position accurately. The experimental results show that when the predicted position and labeling position overlap rate of not less than 0.9, the location accuracy on the test set is 93.3%.2. Through information fusion of vehicle location and classification, the vehicle positioning method based on the multi-task deep learning is proposed. We modified the current three famous deep neural network structures, which combined with SmoothL1 function and softmax function, and chosen the highest classification accuracy of MTGooGleNet to carry out the actual intersection test. The results show that the test set positioning accuracy is 96.4%.3. This paper puts forward a cascade multi-task deep learning model to reduce the influence of the complex background and do fine recognition such as car manufacturer, type and year, which can not be classified correctly by traditional vehicle recognition methods. The strategy of image data enhancement and network pre-training also improved the accuracy. The results show that the accuracy of 196 categories on the standard database of vehicle recognition is 86.67%, about 6% higher than that of traditional method.4. In addition, a standard dataset has been built for vehicle detection, which contains forty thousand pieces of labeling information.