Research On The Key Technologies Of Object Recognition For Intelligent Transportation

With the development of human society, the drawbacks of the existing transportation systems have become increasingly prominent. So people want to make transportation systems more efficient, safe by using various aspects of advanced technologies, that is, realize intelligent transportation systems. As the development direction of future transport systems, the importance of intelligent transportation systems is self-evident. It has a wide impact, ranging from individual’s daily life to the development and prosperity of our country. As an important support technology of intelligent transportation, vision-based object recognition technology is an important foundation for intelligent transportation systems to achieve harmonious relationship among people, vehicles and road.In allusion to the application characteristics of traffic scene, the object recognition framework of this paper is ―motion region extraction first, and then specific object recognition‖. In the recognition of specific object, we select two types of objects — vehicles and pedestrians. In consideration of the adopted framework of object recognition in this paper, we conduct an in-depth study on three key technologies of object recognition in intelligent transportation: optical flow estimation technique, vehicle classification technique, and pedestrian detection and counting technique. Optical flow estimation technique is the key step for motion region extraction in our object recognition framework. Vehicle classification technique and pedestrian detection and counting technique are the recognition for specific objects, and they have many important applications in intelligent transportation systems.The main contributions and innovations of this paper are summarized as follows:1. A fast 3D gradient based optical flow estimation algorithm is proposed. Most of the existing algorithms are too slow to estimate optical flow in real time, for the sake of real time estimation, this new algorithm provides a fast and simple way to estimate optical flow based on 3D gradient computation and plane gradient computation. The main idea is as follow: we can create a three-dimensional coordinate system whose three axes consist of the two axes of image plane and the time axes, so image sequence can be regarded as a cube in this three-dimensional coordinate system. Positions of a point in the image at different times will form a track in the coordinate system, and we can approximate the track curve between two frames with the track curve‘s tangent. So that the point‘s motion between two frames is the projecting line of the tangent of the trajectory on image plane, that is the optical flow. In order to compute trajectory tangent, the main two steps are 3D gradient computation and plane gradient computation, and gradient computation mainly is convolution operations. It can be found that our method can significantly reduce the amount of computation. Compared with the existing methods, the proposed algorithm is simple for computing(mainly a two-step convolution operation) and non-iterative, and the estimation results is available for many real time applications.2. A vehicle classification approach based on histogram features of edge direction is proposed. In this approach, edge direction of edge pixels is used to describe the outline and structure of vehicles, and the edge image is obtained through Canny operator. After edge direction is known, our vehicle classification process is as follow. First, discrete edge direction of edge pixels and compute the histogram of edge direction; Then, using a similar feature organization way as Hog descriptor, organize the histogram to generate feature vector for classification; Finally, SVM is used for learning and classification. Considering the edge pixels are sparse and the purpose of reducing dimension, we make an improvement on the organization of feature vector which is able to reduce the dimension of feature vector. Experimental results show that the proposed feature has comparable classification accuracy with Hog. However, our method is faster as the dimension of the proposed feature is lower.3. A rapid pedestrian detection method is proposed. Our pedestrian detection framework is based on an effective pedestrian detector FPDW. In order to solve difficulties brought about by high density of pedestrian crowd in traffic scene, we first detect both full body and upper body, and then fuse two kinds of detection results to obtain the final detection results. The fusion of the complementary results from two detectors reduces missing rate effectively. In order to meet the requirements of real time applications, GPU is used for acceleration. Experimental results show that the proposed method achieves good performance on both detection speed and detection accuracy.4. A pedestrian counting method based on mixed features and ELM is proposed. The proposed algorithm combines two main pedestrian counting strategies(direct approach and indirect approach). Through mixed features, the information from both direct approach and indirect approach is used in our algorithm, so we can take full advantage of two counting strategies. In order to make full use of the information in pedestrian detection results, we design rLBP feature based on LBP which is able to describe the information effectively. Meanwhile, a relatively new machine learning algorithm – ELM – is used to learn the mapping function between mixed features and the number of pedestrians. A large number of experiments show that our method achieves a significant performance improvement on pedestrian counting accuracy.