Research On Vehicle Detection Method Based On Deep Learning

With the rapid growth of the global economy, car ownership in the whole world keeps increasing, which causes a series of traffic and environmental problems. To address these problems, many countries have begun to study Intelligent Transportation System. The vehicle detection which is a basic aspect of Intelligent Transportation System(ITS) has important implications for developing ITS.The performance of traditional vehicle detection methods such as inter-frame difference method is often not satisfactory. Convolutional Neural Networks which is one method of deep learning is rotationally invariant and shift-invariant. It exhibits excellent performance in many image classification and detection applications. This article will use it for vehicle detection research, mainly to complete the work in the following aspects:1)Positive sample set includes a variety of vehicle with different types, in different appearance or taken from different angles,while negative sample set contains pictures without vehicles. In order to increase the adaptability of the classifier for different lighting conditions and increase the diversity of the sample set, the original samples have a Value shift operation in HSV space.2) Different from traditional vehicle classification algorithms specifying features artificially, Convolutional Neural Networks can extract image features automatically by training the network parameters due to its special structure. We design a vehicle classification model with Convolutional Neural Networks, and train it through back-propagation algorithm. After that, the model with a combination of the detection boxes merging algorithm, applied to the actual scene for vehicle detecting.3)To solve this problem that the size difference between vehicles in near and far places is very large in some scenes for the detection, we propose a multi-size boxes detection strategy. In order to process input images in different sizes, we design a vehicle detection model with Convolution Neural Networks based on Spatial Pyramid Pooling technology. In this model, the features of the whole image were extracted by ZF-5 model. After the feature vectors in uniform size are extracted from Spatial Pyramid Pooling, we chose the SVM classifier to classify them. To solve the problem that a same vehicle is detected repeatedly in actual scene, we use Nonmaxima Suppression algorithm for detection box filtering.Finally, we verify the detection performance of these models in the actual scene, which obtain a better classification results and test results.