Video object tracking using adaptive Kalman filter

Abstract

In this paper, a new video moving object tracking method is proposed. In initialization, a moving object selected by the user is segmented and the dominant color is extracted from the segmented target. In tracking step, a motion model is constructed to set the system model of adaptive Kalman filter firstly. Then, the dominant color of the moving object in HSI color space will be used as feature to detect the moving object in the consecutive video frames. The detected result is fed back as the measurement of adaptive Kalman filter and the estimate parameters of adaptive Kalman filter are adjusted by occlusion ratio adaptively. The proposed method has the robust ability to track the moving object in the consecutive frames under some kinds of real-world complex situations such as the moving object disappearing totally or partially due to occlusion by other ones, fast moving object, changing lighting, changing the direction and orientation of the moving object, and changing the velocity of moving object suddenly. The proposed method is an efficient video object tracking algorithm.

Introduction

The researches for segmenting, estimating, and tracking semantic objects in video have received great attention for the last few years [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26]. The moving object tracking is an important issue in video system, such as surveillance, sports reporting, video annotation, and traffic management system. However, if the background and moving objects vary dynamically, such situations will be complicated. In video analysis, we have to know the features of moving objects, such as color, texture, and shape, etc., so that the moving object can be detected and be tracked. There are some situations for video in a real world environment, including camera lens is fixed or not, multiple moving objects, rigid object, or non-rigid object, occlusion by the other objects, one or many cameras, full automatic or semi-automatic semantic object tracking, etc. According to the above discussion, to conclude, we will encounter the following three problems for tracking moving object.

• ■ Initial moving object segmenting problem.

• ■ Detection of moving object.

• ■ Tracking the moving target in occlusion.

In the initial moving object segmenting problem, the methods of [5], [9], [14] use the user’s assistance to appoint a moving object as target. This is called semi-automatic system. Or, in the environment of still background, [3], [8] apply the difference information of two consecutive frames to extract the moving objects. For the sake of simplicity, to solve this problem, most of methods assume that they are in the environment of still background in the initial state. Although the semi-automatic system simplifies the initial moving object problem, its applications are still popular in real systems.

About the detection of moving object, most of techniques base on the features of color, texture, shape, edge, motion, and shape. Generally, the color feature is frequently used [9], [14], [17], [20], since the human perception is sensitive to the color. However, the disadvantage of the single color feature for moving object detection is that it can only be used in the target object with uniformly distributed color. In addition, Jang and Choi [2] propose an active model to detect and track the moving object. However, the method [2] with the Kalman filter technique which predicts the detecting range to reduce the computational complexity can not be applied to solve the target in occlusion. Therefore, for tracking the moving target in occlusion, Jang and Choi in paper [1] propose the structural Kalman filter to estimate the motion information under a deteriorating condition as occlusion. The structural Kalman filter is a composite of two types of the Kalman filters: cell Kalman filters and relation Kalman filters. Their method partitions an object into several sub-regions and utilizes the relational information among sub-regions of a moving object. The cell Kalman filter estimates motion information of each sub-region of a target, and the relation Kalman filter is to estimate the relative relationship between two adjacent sub-regions. When a sub-region is judged not to be occluded, the cell Kalman filter of the sub-region is enough to estimate its motion information. If a sub-region is judged to be occluded, the relation Kalman filters of the adjacent sub-region are used to compensate the corrupted estimate of the cell Kalman filter. The compensation weighting is set by the degree of occlusion. The idea of the structural Kalman filter method is good, however, its applications are limited since the structural Kalman filter is very complicated and it is not easy to select the criteria or features to partition an object into sub-regions in the different real-world tracking application. In addition, the degree of occlusion needs the other model to judge. More importantly, a complex system is difficult to be expanded to keep tracking multiple objects.

Additionally, many works have been done by motion model based human recognition and estimating human body postures [21], [22], [23], [24], [25]. In general, the above methods offer resistance for the tracker to cope with partial occlusions, changing light condition, and object pose. However, these methods only adapt to human tracking, and more importantly, their computation for model estimation is expensive and the number of model parameters is usually large. And, these methods handle partial occlusions only, failing for severe or complete occlusions. Furthermore, according to the suggestions of the paper [17], for a large range of tracking applications where the object motion is rigid or the object is sufficiently distant from the camera, the tracking system will not need to separate model for the object shape since it can be determined from the object motion. And, tracking on object appearance rather than geometry is easier due to better identification power of appearance features.

In the paper [17], H.T. Nguyen and A.W.M. Smeulders think that the tracking algorithm needs to satisfy the two qualities: simplicity and robustness. Simplicity implies that the algorithm is easy to implement and has the minimum number parameters. Robustness implies the ability of the algorithm to track objects under different conditions which include: severe occlusions, lighting changes, object orientation changing, background clutter, a moving camera, zooms, and rotations. Therefore, in this paper, a simple and efficient method for tracking object using adaptive Kalman filter model is developed and its the robustness quality can be applied to a real-world environment. In the following section, the proposed method is shown. The experimental results are demonstrated in Section 3. Finally, the Section 4 brings to the conclusion and discussion.