Traditional target Tracking Method Introduction _ Traditional target tracking

Because the individual has done a target tracking algorithm before, so it is necessary to do a comb before the work.
Directory
Moving target detection based on the first idea
1. Static background:
2. Sports Ground
Target tracking:
Similarity metric algorithm:
Core Search algorithm:
Kalman Filter:
Particle filter:
Meanshift algorithm:
Camshift algorithm:
Target tracking classification: tracking based on active contour. Feature-based tracking. Region-based tracking. Model-based tracing.

Moving target tracking mainstream algorithms are roughly classified:
Two ways of thinking:
1, independent and prior knowledge, directly from the image sequence detection of moving targets, and target recognition, and ultimately tracking the interest of the moving target;
2, depending on the prior knowledge, firstly modeling the moving object, and then finding the matching moving target in the image sequence in real time.
Moving target detection based on the first idea
For target tracking that does not rely on prior knowledge, motion detection is the first step to realize the tracking, motion detection is to extract the change region from the background image from the sequence image, and the moving target detection algorithm can be divided into motion detection and motion detection under the static background in the context of the relationship between the target and the camera.
1. Static background:
1.1, Background difference: the background of the illumination changes, noise interference and periodic motion modeling
1.2, Frame Difference
1.3, GMM
1.4, optical flow
2. Sports Ground
There are two kinds of cases, one is camera fixed, but the angle of view changes. The second is that the camera is sporty.
Therefore, the first task of detecting moving objects is to estimate and compensate the global motion of the image.
The problem of global motion estimation can be summed up as the estimation of global motion model parameters, and the representation method is block matching and optical flow estimation.
Block estimation: The image is segmented into different image blocks, assuming that the motion vectors on the same image block are the same, and the optimal motion vector estimation is obtained by pixel search.
Optical Flow method: It is assumed that the change of the gray distribution of the image is caused by the motion of the target or scene, that is, the gray level of the target and the scene does not change with time. This makes the optical flow method has a poor ability to resist noise, and its application scope is generally limited and the target and the gray level of the scene remain unchanged under the assumption condition.
Deficiencies: Computational complexity, real-time difficulties.
Target tracking:
The tracking of moving target, i.e., through the effective expression of the target, the process of finding the most similar candidate target position in the image sequence with the target template. To put it simply, it is to locate the target in the sequence image. The effective expression of the moving target is not only modeled on the moving object, the target characteristic expression commonly used in target tracking mainly includes the visual feature (image edge, contour, shape, texture, region), statistic characteristic (histogram, various moment characteristic), transform coefficient characteristic (Fourier descriptor, autoregressive model), Algebraic features (singular value decomposition of image matrices) and so on. In addition to a single feature, the tracking reliability can be improved by fusing multiple features.
Similarity metric algorithm:
After the feature extraction of moving target, a certain similarity measure algorithm is used to match the frame image to realize the tracking. Common methods of similarity measurement include European distance, block distance, chessboard distance and so on.
Core Search algorithm:
In the target tracking process, to find the best matching information and to deal with a large amount of redundant information directly to all the content in the scene, it is very important to estimate the position state of the target in the future with a certain search algorithm, and to reduce the target search scope. One of the most common methods is to predict the possible position of the next frame of the moving object and search for the most advantage in the relevant region. The common prediction algorithms include Kalman filter, extended Kalman filter and particle filter.
Kalman Filter:
A linear minimum variance estimation algorithm for the state sequence of a dynamic system is presented. The method: state equation + Observation equation = Dynamic system, based on the previous state sequence of the system to do the optimal estimation of the next state, the prediction has unbiased, stable and optimal characteristics, and the calculation of small and real-time calculation and so on.
Restrictions: Only for linear and Gaussian distributed systems.
Particle filter:
An optimal algorithm is given to Monte Carlo and Bayesian estimation theory, which is used to sequentially process the measured data in a recursive way, so it is not necessary to store and reprocessing the previous measurement data, thus saving a lot of storage space.
Advantages: Particle filter has excellent robustness in tracking multiple forms of target and nonlinear motion and measurement models.
Mean Drift algorithm (meanshift), continuous adaptive mean drift algorithm (Camshift) and confidence region algorithm are all using the method of parameter estimation to optimize the iterative convergence of target template and candidate target distance, so as to narrow the search scope. (In short, the algorithm is to reduce the search scope is to optimize the search direction)
Meanshift algorithm:
The gradient optimization method is used to realize the fast target location, which can track the non rigid target in real time, and has a good applicability to the deformation and rotation of the target.
Disadvantage: No use of the target in the space of the direction of movement and velocity information, when there is interference around (light and occlusion, etc.), easy to lose the target.
Camshift algorithm:
The extended Meanshift algorithm combines the target color information to improve the mean drift algorithm. The histogram of target image is the probability of color appearing, this method is not affected by the change of target shape, it can effectively solve the problem of target deformation and partial occlusion, and it has high efficiency, but the algorithm needs to manually specify the tracking target before the start.
Target Tracking Category:
Moving target tracking algorithm based on motion expression and similarity measurement can be divided into four categories: tracking with active contour, feature based tracking, region based tracking and model-based tracking. The precision and robustness of the tracking algorithm depend largely on the expression of the moving object and the definition of similarity measure, and the real-time performance of the tracking algorithm depends on the matching search strategy and the filtering prediction algorithm.
1, based on the active contour tracking:
Kass the active contour model, the variable curve is defined in the image domain, and the dynamic contour adjusts its shape to the target contour gradually by minimizing its energy function. Advantages: The gray information of the image and the geometrical information of the whole contour enhance the reliability of the tracking. Disadvantage: The calculation of large, for fast moving objects or large deformation of the situation, tracking effect is not ideal.
2, feature-based tracking:
The overall characteristics of the moving object are not considered, and only some salient features of the target image are followed. (assuming that a moving target can be expressed by a unique set of features, to search the feature set, it is believed that the tracking of the moving target, in addition to the tracking with a single feature, can also be used as tracking characteristics of the feature-based tracking, including feature extraction and feature matching two aspects.
Feature extraction: Extracting features from the original image of the scene.
Feature matching: To match the target feature between frames, and to track the target with optimal matching. The common tracking algorithm based on feature matching is tracking based on two-valued target image matching, tracking based on edge feature matching or corner feature matching, tracking based on target color feature matching, and so on.
Disadvantage: The characteristic robustness is not strong, the noise, the image itself blurred and so on more sensitive.
3, based on the region tracking:
Get the template that contains the target, the template can be obtained by the advance artificial, the template is usually slightly larger than the target rectangle, also can be regarded as irregular shape. In the image sequence, the correlation algorithm is used to track the target.
4. model-based Tracking:
A priori knowledge is established to model the target, then the model is updated in real time by matching the tracking target. For the rigid object, the motion state transformation is mainly translation, rotation and so on, which is convenient for tracking. But for the non rigid object, the geometrical model is not easy to get.
Advantage: This kind of method is not affected by the observation angle of view, has the strong robustness, the model matches the tracking precision is high, suits the maneuvering target various movement changes, the anti-interference ability is strong,
Disadvantages: Computational complexity, slow operation, model update is more complex, real-time poor.

Based on the above ideas, I am using a feature based tracking, why use this? In the final analysis, it is in order to use CNN's powerful Non-linear feature extraction capabilities. Nearly two years, the extensive application of depth learning has brought about a technological revolution in the field of computer vision, the traditional feature extraction method is basically killed by seconds, and using CNN to extract the characteristics of the target, the target can still show strong characteristic robustness under the very complicated conditions such as posture, background, illumination and so on. This is the strength of CNN. The problem of traditional target detection is considered as a classification problem by CNN. For specific targets, direct training of specific detectors, using the TLD framework is still a good way, of course, based on CNN target tracking also has the biggest problem is computing resources, if for some high-definition video streaming processing , the resolution of the image still brings the bottleneck to the performance, but in the engineering application, the traditional method's thought still has the very good reference place, solves this question is not difficult.
Although the individual does not specialize in this direction, however, due to the project needs, or let me have a lot of harvest. Not only the algorithm, but also in the field of engineering applications, there are still many places to learn.
Shortage, welcome to the brick ...