Moving Targets Detecting And Recognizing Technology For Video Surveillance

With the rapid development of computer network and communication technology, the substitution of intelligent monitoring for traditional manual monitoring has become a hotspot in security field. Using intelligent video surveillance (IVS) system to analyze video surveillance sequences intelligently without human intervention has been the development trend in this area.This thesis mainly researches on moving targets detection and recognition combined with video stabilization, and discusses around the practical application of IVS system.Video stabilization technology is used to eliminate physical motion blur and thus eliminate interference for further intelligent analysis and processing. In our work, electrical video stabilization is discussed in detail and existing algorithms are also analyzed. Assuming global jitter vector is consistent, we regard the entire image as a block area on the basis of block matching algorithm, estimate jitter by using gradient difference and then optimize the algorithm. This method can effectively eliminate jitter of videos and is less affected by noises. It can also achieve sub-pixel accuracy and realize real time video stabilization.Based on video stabilization, detection of moving targets in the background of stability is also studied in our work. The modeling speed and model quality are developed and improved for Gaussian mixture model. We use decaying learning rate to improve the updating rate at the beginning of modeling, and suppress update after the modeling stabilization to avoid mixing in those targets moving slowly. Shadow detection algorithms are also used to eliminate shadow, reducing the impact of illumination, improving the stability of background models.Machine learning (ML) method is used to recognize specific targets, while features for ML are studied as focus. We classify common features of images into color features, edge features and texture features. For safety helmet wearing detection, we propose histogram of block-based local binary pattern (HOB-LBP), and use defonnable part model (DPM) as the features carrier. We add histogram of oriented gradient, color frequency, color moment and HOB-LBP into DPM, then use support vector machine (SVM) for training and detecting, therefore achieve safety helmet wearing detection algorithms with a wider applicability and a higher detection rate.In the end of this thesis, we introduce an application of safety helmet wearing detection, and design a safety helmet wearing detection system upon this, including video stabilization, moving target detection and specific target recognition. Coordinative optimization is proposed between video stabilization and moving targets detection modules. A stable Gaussian mixture model background is used as reference frame for video stabilization, and those detected motion regions are eliminated from stabilization computation regions. Meanwhile, motion regions detected by moving targets detection are used as initial detection regions of specific target recognition, thus reducing calculation and improving computation efficiency. And in specific target recognition, we combine pedestrian detection with safety helmet wearing detection and correct output scores of SVM with the inter-frame information, resulting in a higher detection rate and a lower false positive rate.