The application of intelligent visual analysis technology will bring about a considerable market in the future.

[CPS Zhongan Network cps.com.cn] Popularly speaking, intelligent vision technology is to automatically analyze and process the video information collected by the camera, capture the region of interest and the target from the video sequence, and further obtain the time, track, color and many other information of the target, through the analysis of the above-mentioned information of each target.



The core of intelligent vision technology is moving object detection, classification, tracking and recognition technology. Motion target detection is to detect interested objects (such as vehicles or people) in the video image sequence for the use of subsequent steps; the purpose of target detection is to extract the corresponding cluster points from the motion region obtained by motion detection accurately.



_Researchers have developed a variety of moving object detection methods, including inter-frame difference method, optical flow method and background subtraction algorithm. According to the different requirements of practical application, different detection algorithms are compromised among reliability, real-time and accuracy. The purpose of object classification is to extract the moving regions corresponding to human beings from the detected moving regions. Different moving regions may correspond to different moving objects, such as pedestrians, vehicles and other moving objects such as birds, streaming clouds and swaying branches in the sequence images captured by surveillance cameras on traffic roads. It is absolutely necessary for tracking and behavior analysis to classify moving objects correctly. Note that this step may be unnecessary in some cases (for example, when only human motion is known in the scene). Motion target tracking is how to judge the target entering a specific area and track the target trajectory in an environment that needs monitoring.



_can be divided into two situations: one is target tracking in static background; the other is target tracking in dynamic background. Target tracking methods in static background can be divided into single target tracking and multi-target tracking. Target tracking in the static background of a single target means that the camera is fixed in a certain position and the field of view it observes is static. Multi-target tracking refers to multi-target tracking in static environment. It needs to determine the characteristics, position, direction of motion, speed and other information of each target. Target tracking in dynamic background refers to the rotation of the camera under the control of the platform, which can change the image he captured. Therefore, for the whole target tracking process, the background is changed, and the target is moving in the whole process, so tracking is more difficult. Behavior understanding and description is a research hotspot which has been paid more and more attention. It refers to the analysis and recognition of human motion patterns and the description in natural language. Behavior understanding can be simply regarded as the classification of time-varying data, i.e. matching test sequences with pre-calibrated reference sequences representing typical behaviors.



_Current intelligent vision technology mainly focuses on the processing of RGB-based optical images. It detects and tracks targets according to various color spaces, texture structure analysis of targets, gray-scale features and motion features. However, because RGB images can not obtain the distance information of objects in three-dimensional space, the related algorithms are affected by the surrounding environment, illumination changes, background and other factors. Because of its poor robustness in target detection and tracking, it is difficult to realize any target detection and online tracking in complex scenes, and its application is limited greatly.



_Because the image captured by the color camera only retains the two-dimensional information of the space, these methods are all processed in two-dimensional space, only the two-dimensional motion information of the target can be obtained. In the real world, the object is moving in three-dimensional space. If only two-dimensional motion information is obtained, it will be difficult to meet the robustness requirements of practical application. Therefore, more and more researchers begin to use depth images for target detection, tracking and recognition. At present, there are two main ways to obtain three-dimensional information through image sensors. One of the principles is similar to human visual system. Two color cameras are used to shoot the same scene at the same time. The distance between the pixels in the scene and the camera, i.e. Depth, is derived from the parallax of the two cameras. This approach is often referred to as Binocular StereoVision. To derive depth information from binocular images, it is necessary to first calculate the corresponding relationship between the two image pixels, that is, which pixel of the left image and the right image is the same point in the described space. According to the projection relationship, the coordinate deviation of projecting distant points to two image planes is small. Using this principle, the depth can be deduced by coordinate deviation.



However, it is an ill-posed problem to find the corresponding relationship between left and right image pixels, because in essence, the surface captured by color camera does not contain depth information, and the corresponding relationship can only be deduced according to the similarity of appearance, rather than the appearance of the same location may also be similar. Therefore, although scholars have tried various possible means in this direction for many years, they still can not achieve satisfactory results. In addition, calculating the corresponding relationship of pixels and optimizing the depth according to the corresponding relationship and using the correlation between pixels require a lot of computing resources. So far, this problem is still an open one, and it has not been widely applied in production.



Another principle of acquiring three-dimensional information by using image sensors is similar to that of radar: the active method is used to acquire three-dimensional information by using image sensors.