Vision-based detection of moving vehicles using wavelet modulus history images

To improve the robustness and accuracy of vehicle detection, a new detection method of moving vehicles was put forward based on background subtraction and wavelet decomposition modulus history images. First, wavelet decomposition was performed on the original image. In low frequency component, the Gaussian mixture model was used in conjunction with textural features to adaptively update the background image and label initial regions of moving objects. The high frequency component was used to calculate modulus value and obtain modulus history image through history frame accumulation. In view of the fact that vehicle objects have richer edge details than shadow regions, the edges were projected to x and y axes after object slant correction. Using the projection curves, the edge information was iteratively integrated with initial object regions to get final detection results. Experiment results show that compared with commonly used adaptive background extraction methods, the proposed method could detect vehicle objects accurately in practical traffic applications with a capture rate of 99.0% and an effective rate of 92.5%, and could effectively process the object conglutination caused by shadow with higher accuracy.