Research On Key Techniques Of Traffic Monitoring Based On Aerial Remote Sensing

In China, traffic accidents, congestion and other traffic problems have become increasingly frequent in recent years. Traffic problems have become the "bottleneck" of the urban development of economy and society, it is thus very important to solve the traffic problems. One of the generally accepted ways is constructing the Intelligent Transportation System (ITS), which uses new & high technology to transform and manage the existing transportation system, and establishes efficient and convenient road network management system to meet the increasing traffic needs. Traffic information collection system is critical to ITS implementation. Considering the high cost and limited traffic monitoring range of the grounded sensors, it is urgent to develop the wide-ranging traffic monitoring technology. Remote Sensing (RS) with its all-weather capability, large field of view and non-contact observation, makes the wide-ranging traffic monitoring possible. Compared to space remote sensing, aerial remote sensing has higher mobility and resolution, and has been widely used in traffic monitoring.This paper focuses on the key technology of aerial remote sensing in traffic monitoring. Firstly, we investigated the aerial image scene classification method, since the image classification results can provide semantic constraints on traffic monitoring and is the key to improve monitoring efficiency. Secondly, to guarantee the consistency of spatial coordinate systems for different aerial image vehicle detections and the accuracy of traffic monitoring, we developed a new aerial image registration method for moving target monitoring. Furthermore, we proposed a new vehicle detection and orientation method based on Histogram of oriented Gradients (HoG)-based Support Vector Machine (SVM) and deep learning technology. Finally, multi-scale image analysis technology was introduced, and we studied the aerial image fast vehicle detection method. We also developed a vehicle tracking method based on weighted bipartite graph matching. Specifically, the research work and innovations in this paper are mainly the following aspects:1. To solve the problems of hand-crafted feature, we developed an self-learning aerial image classification approach. In contrast with hand-crafted feature, the proposed method extracts local features by unsupervised self-learning and does not need priori knowledge as guidance; then aerial image classification was realized using convolutional feature extraction. Compared to object-based image classification using hand-crafted feature, the training of the proposed method was simpler, and the classification results were better for complex aerial image scene, the Kappa of which increased by 6%.2. To eliminate the influence of projection difference on aerial image registration in object monitoring, a new image registration method was proposed based on Bayesian decision theory. Distribution rule of projection difference was derived and its impact on image registration was analyzed. A large number of feature matches were extracted to calculate the distribution parameters within the pre-set training sample area. The distribution rule was tested with aerial image dataset, and the registered images before and after eliminating projection difference were compared. The difference image results show an improved visual effect and 10% entropy decrease after elimination, which proves that the proposed algorithm is effective in improving the accuracy of aerial image registration.3. A new adaptive method of detecting vehicles from aerial image sequences was proposed in this paper, with the restriction of road obtained from Geographic Information System (GIS). Vehicle detection was executed by virtue of an optimal space searching algorithm in a multi-dimensional space which includes scale, angle and 2D image plane dimension. This paper adopted a typical, commonly used HoG based SVM classifier, and then gained the best responses of the locations and directions of vehicles through specially designed multi-dimensional searching and removal of repetitive responses. This new approach was tested with aerial image sequences of two areas with contrast traffic conditions. The result reveals that the proposed method performs better than unrestricted detection method. Although most efforts have been paid on the development of advanced classifier, the result demonstrates that the detector could also have played a considerable role in vehicle detection.4. A new vehicle detection method from aerial images based on deep learning was proposed. First, a vehicle identification model was constructed based on deep convolutional neural network which takes the original spectral values of image as input, and was trained using labeled samples to learn the model parameters; second, this paper described the vehicle orientation as a regression problem, and introduced transfer learning technology in the model training process, transferred the feature extraction method learned by the vehicle identification model to the vehicle orientation model, then trained the model using labeled vehicle samples. The result revealed that the area under ROC curve (AUC) of the proposed vehicle identification method reached 0.99, and vehicle directional accuracy of the proposed vehicle orientation model was better than HoG+SVM-based multi-dimensional orientation method, and the vehicle orienting process was very efficient.5. A multi-scale image analysis-based fast aerial image vehicle detection method was proposed in this paper. Using multi-scale image segmentation technology the aerial image was divided into several meaningful polygons, and then stepwise refinement method was adopted to achieve fast vehicle detection. the vehicle tracking was described as a bipartite graph matching problem, and three types of weight (distance, angle, similarity) were defined to build the bipartite graph weight matrix, and then Hungarian algorithm was used to obtain the vehicle tracking trajectories. The result revealed that the proposed detection method had multiple advantages over the sliding window detection method. It improved the detection efficiency significantly. The track completeness factor (TCF) of weighted bipartite graph matching-based vehicle tracking method reached 0.96, and the track error (TE) was 3.13 pixels. The vehicle tracking results could be used for macro and micro traffic condition analyses.