The key technology of smart camera

At present, the composition and hardware technology of smart cameras have been relatively stable and mature. In order to eventually complete the monitoring tasks and intelligent technologies of smart cameras, the close cooperation of software functions is also required. Efficient video codec technology and effective computer vision algorithms are the core of smart cameras. Technology provides an important technical guarantee for the camera to perform intelligent analysis tasks. As shown in Figure 1, the structured output from video capture to intelligent results mainly includes steps such as moving target extraction, moving target tracking, moving target classification, and moving target behavior analysis, and structured description.

Figure 1 Smart Camera Analysis Process

1. Sports target extraction

The motion object extraction is the preparation work of intelligent analysis. Based on this work camera, the change region can be extracted from the background region from the image sequence. The effective extraction of the moving target will greatly reduce the computational complexity of the subsequent process. Behavior analysis is of great significance. At present, the mainstream methods include background subtraction, time difference method and optical flow method. The most classical global optical flow field calculation methods are LK (Lueas & Kanada) method and HS (Hom & Schunck) method.

2. Tracking of moving targets

The tracking of the moving target, that is, through the effective expression of the target, finds the position of the candidate target region most similar to the target template in the image sequence. Simply put, it is to target the target in the sequence image. In addition to the effective expression of the moving target, in addition to the modeling of the moving target, the expression of the target characteristics commonly used in target tracking mainly includes: visual features (image edges, outlines, shapes, textures, regions), statistical features (histograms, various moment features). ), transform coefficient characteristics (Fourier descriptive, autoregressive models), algebraic features (singular value decomposition of image matrices), etc. In addition to using a single feature, the reliability of tracking can also be improved by fusing multiple features. The current mainstream methods are: region-based matching tracking algorithm, contour-based matching tracking algorithm, and feature-based matching tracking algorithm.

3. Sports target classification

The classification of moving objects, as the name implies, extracts specific types of objects from the detected motion areas, such as different targets in the classified scenes such as people, vehicles, and crowds. The current mainstream methods include classification based on motion characteristics and classification based on shape information.

4. Analysis of moving target behavior

Behavior analysis is one of the key goals of smart cameras, and it is also a key and difficult issue for video surveillance in maintaining public safety. Behavior analysis involves computer vision, pattern recognition, artificial intelligence and other fields. It is based on the low-level processing of video image sequences. It analyzes and processes the images and videos of the monitoring scene, acquires the information of the monitoring scene or the information of the moving target in the scene, and further studies the nature of each target in the image and the mutual Contact, which leads to the interpretation of objective scenarios and high-level semantic descriptions, often through the use of neural networks and decision trees for behavior analysis.