Research On Vehicle Detection And Tracking Method Based On Video Image Processing

The video detection technology is a very active branch in the research field ofmachine vision. The moving target detection and tracking has a wide range ofapplications in many aspects of national economic and military fields, and scientistsfrom various countries are of great concern to the research in this area.Meanwhile,with the importance of intelligent transportation systems based oncomputer vision technology increasingly significant, the moving vehicle detectionand tracking in the stationary camera image sequence has become an importantresearch area, and played a key connecting role. This paper studys on some key issuesin the moving target detection, tracking and recognition such as moving targetsegmentation, shadow removal, tracking and occlusion based on the theories andmethods of image processing, and applys the research results into practice intelligenttransportation systems. The main research work including the following aspects:1. Shadow elimination methodTraditional methods tend to remove the moving shadows in motion recognition,however, the cast shadows of tall buildings and on-street light as well as changes inthe weather exist in the complex urban traffic environment and the light conditionchange in these scenes will affect the motion detection algorithm properties.Removing illumination changes with the intrinsic image before the motion detectionwill help to achieve robust video surveillance. Therefore, we proposed the shadowelimination method in the pre-processing stage based on the idea of extractingintrinsic image. The method uses the chromaticity characteristics as a measure ofcolor invariance. Firstly, in each color channel, the Gaussian smoothing is used toremove noise, then the geometric mean chromaticity mapping to the logarithmchromaticity ratio space,and we have a series of grayscale images in the feature space by projection. By computing the entropy of the projection data to determine theillumination invariant angle, then come to the natural images with shadow removed,and the images only reflect the physical material changes in the reflectivity and havenothing to do with illumination changes. The final adjustment by the non-lineardynamic range is to enhance the brightness and contrast to make the results morecomfortable with human vision, and conducive to the further detecting andtracking.Simulation results show that the proposed method can effectively remove thevehicle moving shadow and cast shadows by the surrounding buildings in urbantraffic environment.2. Foreground extraction methodMoving object detection is an important part of the Intelligent Traffic MonitoringSystem,and there has been a lot of research in this regard. We first start from thethinking of the sparse background modeling to reduce storage space and computingtime of the background modeling, and make simulation of the moving foregroundextraction method based on edge detection. Using the edge as the characteristics ofthe image foreground extraction method has advantages in efficient, robustperformance. This is because the edge information is still able to be detected in a darkenvironment, even at night it can also be used, and the image edge is less susceptibleto changes in light impact.Then we found that traditional methods can not be appliedto the intrinsic image obtained from the logarithm chromaticity ratio feature space, sowe propose a frequency domain extraction method using three stationary wavelettransform image to extract vehicle area which is in the low-frequency, and thenmorphological post-processing is applied. The simulation results show that thismethod can effectively extract the stationary and moving vehicles and pedestrians,and provide useful data for further tracking and classification. 3. Motion tracking with occlusion handlingIn the computer vision when three-dimensional space projection totwo-dimensional imaging will inevitably bring occlusion, and due to the loss of depthinformation making the solve of occlusion on a flat surface to an incorrectly posedambiguity problem. The perceptual grouping theory provide us discriminant rules todivide block regions based on a large number of psychology experiments.Firstly,according to the rigid body motion characteristics of the vehicle, we set up automatictracking with the geometric characteristics of the movement region.The region feature,relative to the point feature, carrys more information such as the shape and size, etc.and provides the consistency of motion tracking. When the motion region fails tomatch block because of occlusion, we extract the color and contour of the vehiclecharacteristics, and in order to make the mathematical description of similarity andclosed rules we use Euclidean distance of pixels in CIE-Lab color space to define thedegree of color similarity matrix, and then measure the difference between the pixelswith spatial position and profile information. Finally, the block area is separated in theguidelines of the Normalized Cut.