Research On Background Modeling Algorithms And System Design Oriented To Real-time Intelligent Video Surveillance

Background modeling (BGM) is a key low-level technology in intelligent video surveillance, which is used to detect motion objects in video scenes. On the one hand, the BGM’s accuracy and robustness directly affect the performance of the high-level intelligent analysis algorithm. On the other hand, it need to be efficient to leave enough time for the subsequent high-level Intelligent analysis algorithm. So, the speed of BGM is an important performance indicator.Although the theory and applications of BGM have been extensively studied by many experts and scholars around the world, there are still many problems. For the real-time intelligent surveillance system, there are trade-offs between accuracy and speed. The choose of a suitable BGM algorithm should take the requirements of speed and high-level Intelligent analysis algorithm into consideration. Therefore, this paper presents two BGM algorithms for different real-time intelligent surveillance system.Firstly, this paper presents a BGM algorithm based on neighborhood feature and grayscale information. The neighborhood feature in this algorithm is a kind of texture feature, obtained by three-fold encoding of the neighboring pixels. Here, we use this feature to build a neighborhood feature based background model, and then obtain a double-layer background model by combining the neighborhood feature based background model and a grayscale background model. For high frequency regions which have rich texture information, the background pixels can be found by using the neighborhood feature based background model. For low frequency region which have less texture information, the background pixels can be found by using the grayscale background model. This strategy shows the advantages of the double-layer background model. The experiment results show that the performance of the proposed algorithm is among the top ones of existing background modeling algorithms, and it has obvious speed advantage compared with the algorithms in same level. This algorithm runs at 68 fps(i5,2.2GHZ) for the video with width 640 and height 480, which basically meets the speed requirement of real-time intelligent surveillance system and leaves enough time to applying some consuming-less intelligent analysis with the detected foreground objects.Secondly, this paper presents an extended VIBE algorithm based on short-time foreground model. This algorithm is designed to improve the computational efficiency of VIBE algorithm in some videos with many foreground objects. By introducing a short-time foreground model, this algorithm reduces the computation time of foreground pixels from an average of 20 to 6. The experiment results show that this algorithm can improve the speed of VIBE algorithm significantly when the video has many foreground objects. In actual surveillance scenes, this algorithm increase the speed of VIBE algorithm by more than 30 percents. This algorithm runs at 250 fps(i5,2.2GHZ) for the actual surveillance video with width 640 and height 360, and the accuracy and robustness is similar with the VIBE algorithm. This algorithm can meet the speed requirement of real-time intelligent surveillance system and leaves more time for subsequent complex intelligent analysis of the detected foreground objects.Finally, for intelligent surveillance project with ZTE, this paper presents a cross-camera intelligent surveillance system oriented to traffic surveillance network, and applies the presented algorithm to the system. The experiments of this system achieved satisfactory results. Currently, the project has been successfully concluded and delivered.