This paper proposes an end-to-end algorithm for multiple small objects tracking in noisy video using a combination of Gaussian mixture based background segmentation along with a Dynamic Bayesian Networks (DBNs) based tracking. Background segmentation is based on an adaptive backgrounding method that models each pixel as a mixture of Gaussians with spatial prior and uses an online approximation to update the model, the spatial prior is constructed for small objects. Furthermore, we create observation model with hidden variable based on multi-cue statistical object model and employ Kalman filter as inference algorithm. Finally, we use linear assignment problem (LAP) algorithm to perform the models matching. The experimental results show the proposed method outperforms competing method, and demonstrate the effectiveness of the proposed method.
Visual multiple objects tracking has received tremendous attention in the video processing community due to its numerous potential applications in important tasks such as video surveillance, human activity analysis, traffic monitoring, and so forth [8] and [15]. In particularly, small visible space objects tracking is one of the key issues in the research of long-range early warning and space debris surveillance [14]. The basic tracking task consists in estimating the trajectory of a moving object by consistently assigning a label over frames considering noisy measurements. Multiple objects tracking for objects whose appearance is distinctive is much easier since it can be solved reasonably well by using multiple independent single-object trackers. Recently, most researchers are mainly focus on the following two problems. On one hand, it is crucial to improve tracking performances in complex environments. Typical problems are partial and complete occlusion [6], multiple objects [4] and [5], multiple cameras [10], nonuniform object motion, nonrigid and articulated objects [3], complex and time-variable background [2]. For example, in occlusion situation, when tracking a specific object, all the other objects can be viewed as background due to their distinct appearance. The tracker must separate the objects and assign them correct labels [8]. On the other hand, accuracy tracking result is required by many applications, such as providing location of small objects.
The traditional tracking algorithm includes two steps: object detection and tracking. Object detection method mainly regards as background segmentation, which extracting object area from image background. Background modeling is a frequently used segmentation algorithm [11]. Literature [7] and [13] used color and histogram to segment object and background [13] presented a feature integration method to enhance the performance of tracking. Segmentation and tracking are two complementary tasks and several previous methods aim at combining MRF segmentation with object tracking/pose estimation [6]. However, most algorithms simplified the detection step, and focused on the tracking method. Recently, many researches focus on applying probability graphical models (PGMs) to tracking [1], [9] and [12]. The application of Bayesian filtering algorithms to PGMs can provide useful tools for multiple-object tracking.
PGMs are able to provide an appropriate theoretical framework where object dynamics and appearance can be combined and the motion estimation problem can be efficiently solved. It can be divided into two general classes, i.e., directed acyclic graphs (DAGs) and undirected graphs (UGs). DAGs such as Hidden Markov Models (HMMs), Bayesian networks (BNs), DBNs, and Kalman filter models (KFMs) have been extensively used for video analysis because of their capability of modeling temporal relationships. These relationships are characterized by directional dependencies in the time evolution. Instead, Markov networks, Boltzmann machines, and loglinear models are widely employed for image analysis and segmentation to describe spatial dependencies between image pixels. Therefore, we usually use DAGs to solve tracking problem.
In summary, our main contributions are as follows: (1) we propose an end-to-end solution for multiple small objects tracking based on Gaussian mixture and Dynamic Bayesian Networks; (2) we introduce spatial prior to traditional GMM method for small object tracking. The new GMM method can reduce false positives effectively; and (3) we construct a DBN with hidden variable to integrate multi-cue in tracking, which can improve tracking performance (as shown in Fig. 3). The rest of the paper is arranged as follows: Section 2 shows the details of the proposed background segmentation algorithm with spatial prior. The DBNs based tracking algorithm is introduced in Section 3. The experiments are showed in Section 4. Finally, we give the conclusion in Section 5.
In this paper, we have proposed a Dynamic Bayesian Networks based tracking algorithm, and applied to small objects tracking. We introduced spatial prior to GMM based background segmentation method to extract foreground components, and build basic probability graphical model. The observation model with hidden variable integrated multi-cue to increasing the efficiency of DBNs. To address linear and Gaussian of dynamic and observation model, we chose Kalman filter as inference algorithm. Then we employed LAP algorithm to perform the models matching. Experiment results and analysis has shown that the proposed algorithm can track small objects effectively. As future work, we aim to develop an accurate tracking algorithm to non-linear motion and complex environment, such as complete occlusions.
