Target tracing Contour Tracking

Profile tracking

Objects with complex shapes cannot be expressed in simple ry. The contour-based method provides a more accurate shape description. 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 target model can be a color histogram, a target edge, or a contour. Contour-based tracking methods can be divided into two types: Shape Matching Method and Contour Tracking Method. The former searches for the target traits in the current frame, and the latter deducts the new position of the initial Contour in the current frame through the state space model or the direct energy minimization function.

1. Shape Matching Method

This method is similar to template-based tracking. You can search for the target profile and related models in the current frame.

Hutenlocher et al. [1993]-Shape Matching Based on Edge expressions and matching Measurement Using Hausdorff distance.

Li et al. [2001]-A shape matching method using the Hausdorff distance is also proposed.

Another method to match the shape is to find the associated Contour in two consecutive frames and establish a contour Association, or contour matching, which is similar to a point matching method. This method uses the Appearance Features of the target.

-Contour detection is generally implemented by background subtraction.

-After the contour is extracted, matching is achieved by the distance between the calculation Target Model and each contour.

-The target model allows density functions (color or edge histogram), contour boundaries, target edges, or combinations of these information.

Kang et al. [2004]-uses color and edge histograms as the target model

Haritaoglu et al. [2000-used the edge information obtained inside the contour to model the target appearance

It is also an idea to use the light stream method for contour matching. For example, Sato and Aggarwal [2004]-calculate the contour trajectory using the Hough Transform

2. Contour Tracking

This method starts contour deduction from the profile position of the previous frame to obtain the profile of the current frame. The premise is that the target profile of the current frame overlaps with that of the previous frame. There are two different implementation methods for this type of method. The state space model is used to model the shape and motion of the contour, or the contour is deducted directly using the contour energy minimization method such as gradient search.

2.1 tracking with the state space model

The target State is defined by the shape and Motion Parameters of the contour. The status is updated when the posterior probability of the contour is maximized.

Terzopoulos and szeliski [1992]-define the target State with the motion of the control point, and use the spring model to model the dynamic nature of the control point. The new profile state is predicted using the Kalman filter.

Isard and Blake [1998]-define the target State using spline shape parameters and affinine motion parameters, and implement State update using particle filter.

MacCormick and blke [2000]-extended the particle filter-based approach of Isard and blke [1998] to multi-object tracking and blocked by exclusive principles. Able to process occlusion between two targets.

Chen et al. [2001]-uses parameterized elliptical sets to represent outlines. Each profile node has an hmm, and the state of each HMM is defined by the point in the normal direction of the Profile Control Point. The probability of state transition of HMM is estimated using jpdaf. The profile status is estimated by Viterbi algorithm [1967.

The preceding state space model-based tracking methods use explicit methods to express outlines, such as parameter splines. This expression cannot deal with changes in the topology, such as splitting and merging the target region. The method of contour delay can solve the problems caused by topology changes.

2.2 direct energy minimization method tracking

The method of contour deduction for tracking and Object Segmentation is similar. Segmentation and tracking are both achieved through greed.AlgorithmOr gradient descent to minimize energy.

Bertalmio et al. [2000]-Deduct the contour in a continuous frame using the constant constraint of the light stream, and use the level set expression to calculate the contour displacement iteratively. Two Energy functions are used, one for Contour Tracking and the other for intensity changes.

Mansouri [2002]-contour deduction is also performed using the constant flow constraint. This algorithm not only calculates the edge optical flow, but also calculates the optical flow vector of each pixel in the target area. Iteration minimizes energy.

Cremers and Schnorr [2003]-used for contour deduction using light streams

Contour deduction can also be performed by extracting consistent information between the internal and external objects of Consecutive Frames. This method requires profile initialization with the previous position in the current frame.

Ronfrad [1994]-defines an energy function for controlling contour deduction based on the ward distance Static Image Model

Yilmaz and Shah [2004]-use the color and texture model around the target boundary to deduce the target contour
Yilmaz et al. [2004]-A Level Set-based shape model is used to model the target traits and changes. This method can solve the occlusion problem in Contour Tracking.

3. Discussion

Profile tracking is usually performed when the overall target area needs to be tracked. The most important advantage of Contour Tracking is its adaptability to processing target shape changes.

The contour can be expressed in different ways. The most common expression is the binary indicator function. The target region is marked as 1, and the non-target region is marked as 0. for a contour-based method, a contour can be an explicit expression (the contour boundary composed of a set of control points) or an implicit expression (a function defined on a grid ). The most common implicit profile expression is Level Set expression.

The target expression of the contour tracking algorithm can be a combination of motion models, appearance models, shape models, or these models. The target model is typically modeled by parameter or non-parameter density functions, and the target traits can be modeled in the form of contour subspaces.

Appearance-based shape expression is widely used in intuitive contour search. For edge-based shape expressions, hausdroff distance is the most widely used measurement method.

Occlusion processing is also an important issue. Generally, there is no explicit solution. The conventional method is to use the assumption of uniform speed or uniform acceleration to calculate the target location. There are also a few ways to explicitly deal with occlusion with forced shape constraints.

Another important issue is the handling of topology changes such as target segmentation and convergence. This can usually be solved by implicit contour expression.

The important factor for distinguishing different Contour Tracking Algorithms is what features they use, how to deal with occlusion, whether training is required, and more. Some algorithms only use contour boundary information for tracking, other algorithms use the information of the complete area in the contour. The latter is usually more stable for noise. Is a qualitative comparison of different Contour Tracking Algorithms.

This article from the csdn blog, reproduced please indicate the source: http://blog.csdn.net/lynphoenix/archive/2011/02/17/6192069.aspx