Moving target detection and tracking notes-graduation Design

Summarization of algorithm for moving target detection and tracking process
Image preprocessing
Several typical noises in digital images are: The Gaussian noise originates from the noise of the electronic circuit and the sensor noise caused by low illumination or high temperature, and the noise of salt and pepper is similar to the particles of pepper and powder which are randomly distributed on the image, mainly by the image cutting or the error caused by the transform domain;
In general, the introduction of the additive random noise, mean filter, median filter, Gaussian filter and other methods to remove noise, improve the signal-to-noise ratio. The mean filter can be used in the case that the noise distribution is more average and the peak value is not high. Median filter has good effect on the filtering of sharp impulse noise, and it can highlight edge and detail of image; Gaussian filter has good effect on filtering Gaussian white noise.


Moving target detection
Background difference method: The moving object can be segmented completely and quickly. Deficiencies are susceptible to light changes, and background updates are key. The camera motion is not applicable.
Optical Flow method: can detect independent movement of the object, can be used for camera motion, but the computation is complex and time-consuming, it is difficult to detect real-time.
Frame Difference Method: The influence of light changes is small, simple and fast, but can not separate the whole moving object, we need to further use the target segmentation algorithm. There are also some improved algorithms that focus on reducing the illumination impact and detecting slow object changes.


Image identification
The function of image identification is to determine whether the object is independent and there are several moving targets in the image.
1 field: Often take the surrounding 4 or 8 pixels as a field.
2) Connected domain: The two-value image of the interconnected 0-pixel set or 1-pixel set is called the connected domain. The 0 pixels that are surrounded by 1 pixels are called holes. A 1-pixel connected domain without a hole is called a single connected component, and a connection containing one or more holes becomes called a multiple-connected component.
3) Mark: After the difference, a frame image may exist multiple connected domains, each disconnected domain corresponds to a target image area, the work of assigning corresponding marking to each target area becomes a mark.
The identification process is roughly: scanning pixels one by one, scanning to 1 of pixels, detecting the pixel values of its domain, if the same is the connected domain, marking the first target, and then looking for the next target in turn.
After all possible targets have been found, you can draw a gate for each target and frame the target. and a multiple target location list is established, and the central position of each target area is found to be stored as a node in the linked list. The division of the gate may divide the same target into two parts, or a wave door that includes two goals that increase or decrease the error in the target data, and therefore whether the current target is a target or a different target, will be done in the subsequent image segmentation.


Image segmentation
Image segmentation is used to separate the combination of target and background or to separate different goals. Image segmentation can not only compress data, reduce storage capacity, but also greatly simplify the subsequent analysis and processing steps.
1) Histogram threshold segmentation method
The gray-scale histogram is a two-dimensional relationship between the pixel of gray level and the gray scale, which reflects the statistical characteristics of the gray distribution of a pair of images. If the foreground object interior gray value distribution is more uniform, the background gray value distribution is also more uniform, this image histogram will have the obvious Shuangfeng, then can choose two peaks between the bottom as the threshold value. Because the histogram does not contain the location information of the target, it should be combined with the content of the image to determine.
2 Maximum inter-class variance threshold Segmentation method
By using the difference between the image target and the background, the gap between the two types of the total gray level is determined and the threshold is divided.
3) Regional Growth method
Refers to merging pixels that are similar to the surrounding attributes into the target area again.
4) edge detection and Contour extraction segmentation method
5) Morphological Segmentation method
The main function is to make the area of the moving target more complete.
The function of corrosion is to eliminate the object boundary point and remove the object less than the structural element. If there is a small connection between the two objects, when the structural elements are large enough, two objects can be separated by the corrosion operation.
The function of the expansion operation is to merge the background points around the image into the object. If two objects are close, then the expansion operation may connect the two objects together. Expansion is useful for filling holes in the image, and one of the easiest applications to expand is to pick up the cracks.
Morphology can also be used in image filtering, enhancement and so on.


Motion trajectory prediction
After the moving target is segmented, the feature of the target should be extracted, and then the target can be tracked by matching features in the next frame image. In order to reduce the matching region of the search feature and improve the real-time performance, this step is added to predict the trajectory of the target. Motion trajectory prediction is also helpful to enhance the robustness of the tracking under occlusion.
1) Linear Predictive algorithm
2 Kalman filter algorithm and its extended algorithm
3) Particle filter algorithm


Target tracking
1) Feature Selection
Gray-scale features, for gray-scale image, pixel gray value is the most basic target image characteristics;
Geometric characteristics, which reflect the geometrical properties of the target, which is related only to the location of the target pixel, and not to its grayscale. The common geometrical features include target circumference, area, flat rate and height.
Statistical characteristics, such as mean and variance of target gray scale, histogram, entropy, moment and contrast of target phase to background;
Transform domain features, including forier, Gabor, wavelet and other transformation domain characteristics;
Color characteristics.
2) Tracking algorithm
A. Template matching tracking
The template slides on the image, corresponds to the gray value of each position of the image, compares with the gray value of the pixel on the template, calculates a cumulative error in each position, picks out the most suitable position and completes the match.
B.camshift Tracking
Color histogram matching.
C. Active contour Line Tracking
Also called Snake algorithm, minimizing the energy function of the image.
D. multimode tracking
Use multiple tracking algorithms (models) to track simultaneously.


The complexity of moving target detection and tracking is shown in
1, light changes. Due to the change in time, morning and afternoon, daylight intensity and angle changes will cause light changes, and due to the weather, the same will lead to light changes. Due to the change of light, it is difficult to make a detection algorithm adapt to various illumination conditions.
2. The interference of moving target in the scene. For example, the movement of various targets in a large area, the sudden stop and sudden start of a vehicle, the frequent movements of certain objects in a scene, such as swaying branches and leaves, fluctuations in the water surface, etc.
3, initialization problem. In some monitoring scenarios, it is difficult to get a pure background image without noise interference (images without detection targets and motion background targets). For example, the busy traffic scenes of people's cars.
4, occlusion and hole problems. To detect the moving target, by the background of the target occlusion, how to judge Occlusion.
5, shadow problem. In the target detection, how to distinguish the target from the shadow that it produces, so that only the target part is detected.
6, the disappearance of the target. The moving target stays in the scene for a long time and may become a background target.
Because of the complexity of target detection, it is not realistic and impossible to establish a universal target detection algorithm which is suitable for all cases. Therefore, according to the concrete situation, it is the research direction of target detection method to establish the target detection algorithm which accords with the actual condition.