IMPROVING THE METHOD OF DETECTION OF DYNAMIC OBJECTS IN VIDEO SEQUENCES

Abstract

Many computer vision programs must first find moving objects in a video frame. As an example of video surveillance, we need to detect an intruder in a monitored scene. To achieve this, the image can be averaged over time during the initialization period. This is done in order to separate moving objects from a static background. The main process used for this task is background subtraction. A wide variety of background subtraction techniques have been proposed over the years, which can be grouped into codebook-based, probabilistic-based, sample-based, subspace-based, compressed sensing-based, and more recently, deep learning-based approaches.

Background subtraction plays an important role in video surveillance systems, as it is one of the most widely used motion detection tools. If scientific progress has made it possible to develop sophisticated equipment for this task, the algorithms used must also be improved. Over the past decade, a background subtraction technique called ViBE has been gaining popularity. However, the original algorithm has two main drawbacks. The first is the ghosting phenomenon, which appears if the initial frame contains a moving object or if the background situation suddenly changes. Second, it doesn't work well on complex backgrounds. This paper presents an efficient ViBE-based background subtraction approach to address these two issues. It is based on an adaptive radius for working with complex backgrounds, a cumulative average and a pixel counting mechanism to quickly eliminate the phenomenon of phantom and adapt to the sudden change of the background model.

The paper presents an efficient background subtraction algorithm based on ViBE for a complex background. The proposed method was explained in detail with all the necessary parameters. It combines advanced ViBE, which uses cumulative averaging and a pixel-counting engine for fast de-ghosting, and adaptive ViBE, which calculates an adaptive radius based on background changes. The obtained results are given, which demonstrate the effectiveness of the method in comparison with the existing ones. With an average frame rate of 30 frames per second, it can also be used in real-time applications.