Real-time Detection And Tracking Of Human Based On Head And Shoulder Feature

From the beginning of human detection and tracking technology, it is always an important research direction of computer vision. Human detection and tracking fusion of computer science, image processing, pattern recognition and many other fields of technology, and because human body target motion is non-regularity and background is complex, human detection and tracking has been a challenging task, so far there is not a perfect solution. In this paper, we do some research on human detection and tracking algorithm in the situation that camera is static and background is not very complex. This paper is divided into three parts: moving object extraction, human target detection and human target tracking .Moving Objects Extraction refers to the image sequence with moving objects will be separated from the background, this paper describes the most commonly used three kinds of moving object extraction method: The background difference method, frame difference and optical flow method. The basic idea of background difference method is using the knowledge of statistics to determine whether the target pixel belongs to the former sites through the establishment of a background model, such as Single-Gaussian background models and Gaussian mixture background model. The basic idea of frame difference method is through the difference of continuous two or three frames , using threshold segmentation, has achieved the purpose of separation of foreground and background. Optical flow method is an object extraction method used in situation of moving camera, because of its high computational complexity it is not suitable for real-time systems. This paper focuses on the three-frames difference method, and introduced the method of selecting, experiments show that inter-frame difference method is more suitable than the background difference method for human detection and tracking technology. The obtained binary image of frame difference is usually not continuous and has hollow, and also there is noise, so we need do restoration and de-noising for the image, in this paper, we use basic opening and closing operation of mathematical morphology to do the restoration and de-noising, get a good result. Finally, this paper describes how to use the three circle method to mark out the connected area of foreground image, each connected area will be handled respectively.So far human target detection has been got a lot of research, its basic idea is to select the appropriate features of the human target, then determine whether the detected objects is the human target through the features. According to different methods of selecting characteristics, human target detection can generally be divided into model-based, statistics-based and image sequence-based, human detection based on head and shoulder feature is belong to model-based human detection, model-based human detection is closest to the true course of human judgement, through the contours of moving objects, shapes, etc. to determine whether they are in accordance with the model of human target. Part of the body's head and shoulders is not easy to be changed relatively, also not easy to be blocked, so it is is a better test characteristics of non-rigid body. This paper describes how to use the body's vertical projection histogram to extract a model of human head and shoulders, this extraction method has a high adaptability, not only can extract the integrity of the normal upright human body, but also can extract the half-length of the human body and the human body in a squatting state. On the aspect of feature extraction of head and shoulder, this paper describes the principal component analysis (PCA) and its improved version of two-dimensional principal component analysis (2DPCA), PCA is a method for dimensionality reduction, drop the original data from high-dimensional space to low-dimensional space, and preserve most of the information of original data, when Used in image processing, PCA generally expand the image to a one-dimensional row vector or column vector, then reduce dimension. Because images to be processed are usually higher resolution, after expanded to one-dimensional the dimension is usually very high, computational complexity is very high, 2DPCA have been proposed in order to solve this problem, 2DPCA has the same basic idea with the PCA, but the calculation does not require the image to expand to a one-dimensional vector, calculated directly as a matrix, greatly reduces the computational complexity, improves computing speed. In the model-based human detection, first we need a standard model to match, as the ever-changing body posture and face a different camera angle, so we need a sample database that contains more than enough head and shoulder samples, this paper makes a sample database contains 800 head and shoulder samples, and uses 2DPCA to extract the feature for the following detection. When detecting, 2DPCA do the image dimension reduction calculation as a matrix, so this paper pulls in the concept of matrix simulation to judge whether the dimension reduced detected image is matching sample feature, finishes the human target detection. When there are many humans and block each other in the video, detection will be very difficult, because when humans block each other the foreground extracted will be a whole binary image, program will treat it as one connected region to detect, then get a wrong result, this paper takes a situation of two humans block each other as an example, presents a horizontal projection histogram-based isolation method, achieved some success.Human target tracking technology can be used on locating and tracking human body targets of interest for a long time, has been widely used in the cross-border alarm and human abnormal behavior recognition. As a reasoning problem, human target tracking usually need mathematical tools to finish the reasoning process. This paper achieved tracking of human targets through Kalman filter, Kalman filter is one of the most optimized autoregressive data processing algorithms, it can estimate the current state of dynamic systems from a range of measurements which contain the noise, and then predicts the next time location of moving target. After using human target detection technology to locate a human target, this paper takes centroid of the smallest rectangle surrounding the human body target as the tracking target, uses position and speed of centroid to initialize Kalman filter, then complete tracking through a range of iterative operation of Kalman filter. Because Kalman filter has a characteristic that it will accumulate error stepwise, when tracking we can use some parameters as the threshold, judges whether the error is too large, if the error is too large we will use the detection program to correct it.Human detection and tracking is quite a new task in computer vision, though now many people and organizations do a lot of research and bring forward many effective methods, to realize a human detection and tracking system that has robustness, veracity and speediness is still a very difficult challenge.