Overview of Motion Target Tracking Algorithms

Moving Target Tracking is an indispensable part of the video monitoring system. In specific scenarios, some classic algorithms can achieve better target tracking performance. This article introduces general target tracking algorithms, compares several common algorithms, and introduces particle filter algorithms and contour-based target tracking algorithms in detail. Finally, we briefly introduce the processing of object occlusion, multi-camera target tracking, and target tracking under camera motion.

I. general target tracking algorithms

Generally, target tracking is divided into two parts: Feature Extraction and target tracking algorithms. The extracted target features can be roughly divided into the following types:

1) The color histogram of the target area is used as the feature. The color feature has rotation immutability and is not affected by changes in the size and shape of the target object. The color space is roughly the same.

2) The contour features of the target are faster than the algorithm, and the results are also better when the target has a small part of occlusion.

3) The texture features of the target are more effective than contour features.

Target Tracking Algorithms can be roughly divided into the following four types:

1) Mean Shift algorithm, I .e., meanshift algorithm. This method can quickly find the most similar position with the target through a small number of iterations, and the effect is quite good. However, it cannot solve the occlusion problem of the target and cannot adapt to the shape and small changes of the moving target. The improved algorithm is camshift, which can adapt to changes in the size and shape of the moving target and has a good tracking effect. However, when the background color is close to the target color, it is easy to increase the target region, which may lead to loss of target tracking.

2) Target Tracking Based on Kalman filter. This method assumes that the motion model of an object complies with the Gaussian Model to Predict the motion state of the object, and then compares it with the observation model, update the state of the Moving Target Based on the error. The accuracy of this algorithm is not very high.

3) Target Tracking Based on particle filter. The distribution of the re-sampled particles is obtained each time based on the current tracking result, and then the particles are spread based on the particle distribution, then, the state of the target is re-observed through the diffusion result, and the status of the target is normalized and updated. This algorithm is characterized by fast tracking and can solve partial occlusion of the target, which is increasingly used in practical engineering applications.

(4) Modeling of moving targets. In this method, we need to know in advance what the target object is, such as vehicles, pedestrians, and faces. By modeling the target to be tracked, and then using the model for actual tracking. This method must know the target object to be tracked in advance, and then trace the specified target. This is its limitation, so its promotion is relatively poor.

Ii. Moving Target Tracking Based on Particle Filter

Particle Filter supports moving object tracking in four steps: 1) initialization and extraction of Moving Object Features; 2) feature search; 3) decision-making; 4) particle re-sampling. The following three processes trace the Moving Target cyclically.

1) Extract moving target features

In this phase, you need to select the target to be tracked through human-computer interaction, and calculate the color space histogram of the region by calculating the features of the target, such as the color features of the target, to describe the features of the target.

2) feature search stage

After extracting the features of the target, you can search for the target object by scattering particles around the target. There are many ways to scatter particles, such as: a) evenly distributed placement: evenly scattering particles in the entire video screen. B) scatter the particles in the vicinity of the target with Gaussian distribution, that is, scatter more near the target and put less farther away from the target. Each particle calculates the color feature of its position based on the Target Feature obtained during initialization, and calculates the similarity between the histogram vector and the target histogram vector. Then, normalize the similarity calculated between each particle and the target so that the similarity of all particles is equal to or equal to 1.

3) decision-making stage

Weighted average is performed based on the similarity between each particle and the target. If the coordinate of the first particle is that it is similar to the target, the most possible position of the target is:, and N is the number of particles.

4) particle re-sampling

Particle re-sampling is to re-allocate the number of particles based on the similarity, that is, place more particles out of the particles with the highest similarity, and place fewer particles in the particles with low similarity. The core idea of the particle filter algorithm is random sampling of particles and heavy sampling of particle importance. Because we don't know where the target is, We randomly scatter the particles. After a particle is sprinkled, the similarity of each particle is calculated based on the feature similarity of the target to be described. Then, more particles are scattered in areas with high similarity, and less particles are scattered in areas with low similarity, this is the idea of particle filtering.

3. contour-based tracking

It is difficult to describe objects with complex shapes with simple geometric shapes. The contour-based target tracking method provides more accurate shape descriptions. The main idea of this method is to use the target model created in the previous frame to find the target region of the current frame. The model can be a color histogram, edge, or contour of the target area. Contour-based target tracking methods can be divided into shape matching and Contour Tracking. The former searches for features of the target in the current frame, and the latter uses the state space model or directly minimizes the energy function to derive the position of the initial Contour in the current frame.

1 , Shape Matching Method

This method is similar to the template-based target tracking method. The basic idea is to search the target profile and target-related models in the current frame. Another method is to find the associated Contour in a continuous frame, and use the Appearance Features of the target to establish the contour association, or contour matching, which is similar to the point matching method.

First, the contour is detected by the background subtraction. Then, after the contour is extracted, the distance between the target model and each contour is calculated to achieve matching; the target model can be a density function (color feature or edge histogram), a target edge, a contour boundary, or a combination of these information.

2 , Profile tracking

The contour of the current frame is predicted from the contour position of the previous frame. There is an overlapping part between the current frame and the target profile of the previous frame, which can be roughly divided into two implementation modes. The state space model is used to model the shape and motion state of the target profile, or use the gradient search to deduct the contour. Gradient search is a method to minimize contour energy.

1) Target Tracking of the state space model

The state of the target is defined by the shape of the contour and the Motion Parameters of the target (such as velocity, acceleration, and motion direction), and the target State is updated when the posterior probability of the contour is the maximum.

2) Target Tracking of the Energy Minimization Method

This method is similar to the target segmentation method, where the segmentation and tracking part are minimized by the gradient descent method or greedy algorithm.

3) Discussion

Contour Tracking is usually used to track the entire target area. Its most prominent advantage is that it can better adapt to the changes of the target shape. In the contour tracking algorithm, the target expression can be a combination of appearance model, shape model, motion model, or these models. The density function of the target model is used for modeling, and the target shape is modeled in the form of contour subspaces. Shape expression based on appearance is widely used in intuitive contour search. Hausdroff distance is the most extensive method for measuring edge shape-based expressions.

Iv. occlusion Processing

Generally, occlusion can be divided into three types: Inter-object occlusion, background occlusion, and self-occlusion. For mutual occlusion between targets, you can choose to deal with this problem based on the prior knowledge of the target location and target features. It is difficult to judge the partial occlusion caused by the scene structure because it is difficult to identify whether the target shape is changed or blocked. Therefore, the general method to deal with occlusion is to use linear or non-linear dynamic modeling to model the moving target, and predict the possible position of the target when the target is blocked, the position of the target will be corrected until it appears again. The kalman filter can be used to estimate the target location, or the particle filter can be used to estimate the target State. Alternatively, you can use contour projection, optical flow features, and other methods.

V. multi-camera Target Tracking

The need for multi-camera target tracking generally comes from the following two aspects:

1) Deep tracking and occlusion processing;

2) Expand the tracking scenario area.

An important issue in multi-camera tracking is how to deal with the ing between different cameras. Generally, it can be manually defined or automatically calculated based on the observed scenario and moving target.

6. Target Tracking under camera motion

When the camera is moving, it cannot obtain the specific position and size description of the target through the background subtraction. In this case, it is highly required to describe the features of the target and the target tracking algorithm. There are several common methods:

Centroid tracking algorithm

This tracking method is mainly used to track bounded targets, such as airplanes, and large differences between targets and backgrounds. When a target is completely in a video image, image preprocessing algorithms, such as contrast enhancement, image denoising, and bipolar enhancement, are often used for tracking.

Edge Tracking Algorithm

When the target to be tracked has one or more edges and has uncertain edges at the same time, Edge tracking is the best. A typical tracking object is a rocket in the launching process. It has a good front edge, but the tail edge is uncertain due to the jet.

Scenario lock Tracking Algorithm

This algorithm is mainly used for target tracking in complex scenarios. Suitable for monitoring on the ground or on the ground from the air. This algorithm will track multiple targets in the video screen, and then estimate the global motion of the entire scene based on the motion status of each vertex. The moving targets and moving targets in the scene are automatically selected. When the target to be tracked is moved out of the video screen, the new target can be automatically identified.

Combined Tracing Algorithm

This method, as its name implies, is a combination of two or more algorithms with complementary characteristics: for example, a combination of centroid algorithms and Related algorithms. The combined algorithm is suitable for scenarios with large changes in target size, shape, and surface features.