Research On Key Technologies Of Traffic Anomaly Video Image Detection Based On The Local Invariant Feature

Abnormal traffic detection, as an important part of intelligent transportation, not only reduces the working intensity of traffic managers, but also improves the efficiency of traffic management. However, the actual traffic environment is usually associated with complicated background, which leads to illumination changes, occlusion, scale and size changes, the change of perspective affine transformation and noise interference factors. Thus, abnormal traffic detection algorithm based on trajectory becomes invalid due to the failure of target extraction and tracking of the intermediate links.To resolve the above issues, we propose an abnormal traffic video image detection technology based on the image local invariant features, which does not need the target detection and tracking, and thus avoids the defects of the existing trajectory based methods. To deal with the difficulty and the insufficiency inherent in the technology, we study some key technologies such as the local invariant features, image representation model and the traffic anomaly detection algorithm. The main results of this work are summarized as follows:1. While image local invariant feature lays the foundation for image description, SIFT feature has good resistance to light changes, occlusion,scale and size changes, the change of perspective affine transformation and the influence of noise. Thus, using SIFT feature for abnormal traffic video image detection can effectively solve the problems caused by the interference of complex traffic environment.Since the traffic video images are color images, the existing SIFT feature extraction and description methods convert color images to gray image or simply extend the feature extraction and description method from grayscale image to color image, not fully exploiting and using color image correlation and complementary information of each channel.Geometric algebra is an efficient tool of mathematical analysis and calculation, the project using the geometric algebra as the mathematical theory framework built geometric algebra representation model and calculation method that contains color domain and spatial information of color image. It proceed the mathematical deduction for the color of color image characteristics of the domain and airspace and put forward the color image geometric algebra to build scale space method. On this basis, we put forward a kind of color image SIFT feature extraction and description based on geometric algebra algorithm.Theoretical analysis and experimental results show that the proposed algorithm is superior to the traditional SIFT algorithm for effectively resisting various transformation appeared in the traffic image affine and noise impact and compression transformation, etc.2. The representation based on the images feature extraction of traffic images is another key technology for the abnormal traffic video image detection. The bag-of-word model has been widely applied in image description and classification,however, the fuzziness and uncertainty of the word map and characteristics in the current bag-of-word model, reduces the accuracy of the image description algorithm based on bag-of-word model. Fuzzy set theory is an efficient mathematical tool to deal with fuzzy and uncertain problems. Therefore, the project puts forward a bag-of-word model framework based on fuzzy set theory. Within this framework, we propose a bag-of-word model based on exponential and Gaussian membership functions. Theoretical analysis and experimental results show that the proposed algorithm effectively improves the accuracy of the bag-of-word model on the image.3. Fuzzy problems are solved by geometric theory and methods in fuzzy geometry which combine the advantages of fuzzy set theory and geometric theory. Considering the fuzzy and uncertain problems in the image description algorithm, the project defines fuzzy characteristics, fuzzy visual words, visual dictionary as well as the fuzzy similarity measure in higher dimensional fuzzy geometric space and build the image representation model based on high dimensional fuzzy geometric. Following this model, we develop the SPM algorithm and LLC algorithm which are all based on fuzzy geometry. Theoretical analysis and experimental results show that the proposed algorithm in the traffic image representation have validity, and has better image accuracy than traditional SPM algorithm and LLC algorithm.4. The performance of traditional abnormal traffic detection algorithm is poor due to the complex background changes, illumination variations, scale and size changes, geometrical changes and obstruction. As a remedy, the image representation algorithm based on local invariant features can effectively resolve these difficulties. Therefore, this paper develops an abnormal traffic detection algorithm based on images representation algorithm under a Gaussian model. First, we adopt the technique of video image block and image invariant features to represent the image, then we propose a traffic parameter to reflect the traffic. Based on this, we use maximum likelihood estimate to establish the normal traffic Gaussian distribution model, and use it for abnormal traffic video images. The experimental results show that the proposed algorithm can effectively detect the intersection of this chapter and the urban main road traffic anomalies. The proposed algorithm does not require target detection, target tracking and forming trajectory. Therefore, it is not affected by target detection and tracking failure. Moreover, the image representation based on the local invariant features of images guarantees that the proposed algorithm is robust to the variations of illumination, scale and size changes, geometrical changes and obstruction, etc.