Adopt differential detection method to detect moving target and perform image processing, code
introduction:
1. The method of moving target detection in video:
1. Pixel-based moving target detection (pixel color value, pixel position)
2. Region-based moving target detection (correlation between pixel attributes, large-scale features)
3. Other methods
2. Due to the complexity of the video environment and the complexity of application scenarios, it is difficult to use a single technology to deal with all application environments, and various related algorithms are constantly emerging.
Problem statement and application:
Moving target detection generally does not constitute an application alone, but as a component of the system.
With different system applications, the specific requirements of general target detection are also different.
1. System Statement
1. A high-level model generally has three modules (detection, tracking, and recognition), of which the recognition module is not necessary.
2. The detection module can be subdivided into two small modules (moving target area detection and moving target grouping).
3. The input of the model is continuous video, and the output is tracked objects or classified objects.
2. Key discussion content: Determine the moving target area
1. How to extract the foreground of movement
2. How to build a good background model (not the main task, but an important auxiliary tool to complete the task goal)
3. Background subtraction
Pixel-based moving target detection:
1. Mean-threshold method
2. Shadow removal and three Gaussian model
The solution to the shadow: Abandon the gray value and use color value modeling.
Hypothesis: The shadow cast by the moving target on the road mainly changes the brightness of the position without a big influence on the chromaticity.
HSV is easy to solve, RGB is slightly more complicated.
Use brightness and chroma deviation values to divide the new colors into four categories: foreground, background, shadow, strong light
A more complex method: the three Gaussian model .
Three Gaussian distributions are used to model roads, moving foregrounds and shadows respectively.
The main problem: how to obtain the parameters of each Gaussian distribution through a certain period of learning to establish an effective model. (EM algorithm, incremental EM algorithm).