Research On Methods For Traffic Incident Recognition Based On Video Processing In Mixed Traffic

Mixed traffic is a main feature of the traffic situation in China. Traffic incident detection based on video image processing is an important research area in intelligent traffic control. This thesis starts with an introduction to video-based detection in mixed traffic, and then focuses the discussion on background model, object identification and incident detection.First of all, in this thesis, an adaptive background model based on mathematical morphology is proposed, which overcomes the difficulty in extracting the background characteristics of a mixed traffic environment with multiple types of moving targets. The proposed model can rule out the interference caused by moving objects in the foreground, so that it can automatically generate and update the background of a complex traffic situation. At the same time, the model can adjust the image brightness in real time corresponding to the ambient light intensity to ensure that different video images have similar background brightness.Second, by recognizing morphological characteristics of the moving objects, we propose a feature extraction and expression method based on eccentric vectors, which has a rotation, translation and scale invariance in motion of objects in mixed traffic. Support Vector Machine multi-classification learning mechanism is applied here to establish an SVM-based algorithm that can effectively identify motor/non-motor vehicles and pedestrians. For motion tracking, we propose a multi-feature matching method based on the Kalman-filter model, which can ensure an accurate estimation of the motion state. This part of our research provides effective technical means for identification of the target incident.Finally, we develop a context-based expression and identification method for traffic incidents, which makes use of state properties of the moving target and incorporates context information to build incident unit, so as to achieve simple common semantic representation of the traffic incidents. As for incident identification, we propose a basic incident identification method based on classification instrument in combination of logic constraints, and verify that our method can effectively identify basic traffic incidents in an experiment for incident identification of pedestrian crossing violation.