Overview of target tracking methods

Monitoring the object tracking problem in video, many scholars put a lot of effort to research, has produced a variety of different tracking methods. For the tracking method, we can classify and describe the visual feature extracted by the algorithm, the method of locating and tracking the target, and the number of the target. The following is a description of the key points in the tracking method.

Tracking the visual characteristics of an object

Choosing the right visual features is essential for tracking algorithms. Generally selected visual features can be used as the target unique description, making it significant in the feature space. It is important to note that the selection of features is closely related to the presentation of the target. For example, based on the common color histogram in the target model of the region, the contour-based algorithm investigates the edge features of the object. Many tracking algorithms employ two or more federated features in real-world applications. Common visual features are:

Color

The appearance color of the object is mainly determined by two factors, one is the spectral power distribution of the luminous body, and the other is the reflective performance of the object surface, which is dominated by the latter in the surveillance video. In image processing, RGB (red, green, blue) color space is usually used to represent the data, but in the RGB space, the dimensions are highly correlated and do not represent the perception of color change well, and the actual use of HSV (hue, saturation, brightness), LUV and other space to represent color.

Edge

Moving objects ' boundaries can cause changes in image pixels. Edge features are less sensitive to changes in light intensity than color features, so algorithms that track target boundaries often use edges as features. For example, the canny edge detector is one of the most common edge features due to its simple algorithm and high accuracy.

Optical flow

Optical flow is a dense field of displacement vectors used to represent the instantaneous velocity of each pixel in a region. The luminance constraint is used to calculate the optical flow, that is, the luminance of the corresponding pixel points in the adjacent frames of the video image is constant. Optical flow method is commonly used in motion-based object segmentation and tracking algorithms.

Local feature descriptors

The feature points extracted from the local area of the video image are often invariant to illumination, rotation or scale, and have strong extensibility. The common features are LBP and sift characteristics.

Location Tracking Target Method

After extracting the target features, the tracking algorithm needs to locate the target in the video image sequence according to certain rules, and the common methods are:

Goal Similarity Metrics

The similarity between the two objects in the time period of the adjacent frame or sliding window is calculated and matched to find the corresponding correlation, in which the Euclidean distance, the Babbitt distance and the checkerboard distance are widely used.

Target Search algorithm

The information contained in the video image contains a lot of redundancy, and it wastes a lot of computational resources if the image is matched directly to the whole picture. The corresponding solution is to predict the location of the possible region according to the target location and the state of motion, and the subsequent video sequences are searched only within the specified range, which includes the potential location of the next frame of the predicted target, including Kalman filter and particle filter. There are also optimization algorithms that reduce the search range by iterative convergence, including mean shift algorithm (MEANSHIFT) and continuous adaptive mean Shift (CAMSHIFT) algorithm.

Probabilistic Tracking method

Using Bayesian filtering theory, probabilistic tracking method solves the problem of state estimation in video tracking field. In this kind of algorithm, the target state is usually described by position, velocity, scale or rotation angle, and the subsequent state of the tracking target is predicted according to the state transfer model, and the execution degree and the predicted value of the model are corrected by the actual observation value. The Hidden Markov model (Hidden Markov model,hmm) is often used in probabilistic tracking algorithms.

Tracking algorithms focus on the number of targets

From the angle of the target quantity of the algorithm, the existing tracking algorithms can be divided into two categories: single target and multi-objective. The single target tracking algorithm focuses on estimating the state of the target based on the appearance and motion clues, and the common algorithm is briefly described above. The multi-target tracking algorithm is concerned with multiple objects in one frame image, which needs to correspond with multiple targets. In the multi-target tracking algorithm, the first problem is to solve the correspondence between the observed object and the target in the video, that is, the data association problem. The data association is related to the object state, once the correspondence of the target in the video is determined, the filter algorithm can be used to estimate the state of the target.

The multi-target tracking algorithm can be directly extended by the single target tracking algorithm, which initializes a number of single target tracking sub (Tracker) to track each target, and initializes a new tracking child each time a new target is detected. Although this kind of method can take advantage of the more mature single-target tracking algorithm, but it does not take into account the impact of each goal, the general improvement method is to introduce the online learning mechanism, the other targets as negative samples to learn, so as to avoid the two target intersection of the ID Exchange error occurred. There is a class of multi-objective tracking algorithm is to abstract the problem into the graph theory of the specific problems, such as the maximum weight of the graph, such an algorithm needs to target detection, the target detection results are generally used as the vertex in the graph, the use of a certain strategy between the target edge, the weight of the edge is generally the similarity between the target. Compared with the multi-target tracking algorithm which is extended by single target tracking, this kind of algorithm comprehensively considers the global information, and uses the more mature graph theory algorithm, has the complete Theory Foundation, the experiment result also shows that it achieves the better multi-target tracking performance.

For the second type of multi-target tracking algorithm described above, the common framework, as shown, is generally referred to as detection-based tracking.

Overview of target tracking methods