The performance of the statistical algorithm is the soul of the video people counter system. The industry's level video passenger counter (video based) can be divided into two categories according to the different core technologies used by the algorithm.
The first is based on foreground object detection technology. It uses background differences between frames to create background models, filter out foreground objects and determine the movement path of objects. But the application of the technology has an important premise: the requirements of the scene are open, open, such as the border line. This is because the object is occluded or crowded, objects of different prospects will be unable to distinguish the conglutination, motion estimation based on the prospect of failure also, to establish the correct crowded scene will also affect the background model, and in the open environment, this situation rarely happen. However, most of the actual statistics do not satisfy the "empty scenario" hypothesis of the technology, so the products using this algorithm have two characteristics.
First, the camera is required to be placed vertically down. This is because this kind of algorithm is very sensitive to the object occlusion.
Second, with the increase of passenger traffic congestion, the accuracy of the number of statistics will be significantly affected. As shown in Figure 1, even if the background model is not affected by congestion, the algorithm based on foreground object detection still fails to distinguish the number of people because of the inability to distinguish the adhesive body.
The accuracy of the statistics will be significantly affected. Even assuming that the background model is not affected by congestion, the algorithm based on foreground object detection still fails to distinguish the number of people because of the inability to distinguish the adhesive body.
The basic principle of the second class algorithm is close to the recognition process of the human visual system, and the Wen'an statistics module is the leader of this kind of algorithm. The statistical algorithm of Wen'an population is divided into two stages: detection and tracking. The number tester has a detector trained by a large number of samples. In the detection stage, the detector can locate each person's location in every frame video. This detector can be separated according to a single frame of human and other objects, is the essential difference between Wen'an statistical algorithm and the first kind of algorithm, which ensures the tracking phase, the number of Wen'an statistics module tracking object is a separate person, and not susceptible to crowd or other objects interference.
Base algorithm based on excellent
Our statistical module has the following characteristics: excellent accuracy. When the other conditions are normal, the accuracy time is 15 minutes or more, and the camera is vertically placed (90 degrees), the accuracy rate is >90%.
Good camera angle adaptability
Under normal conditions of other conditions, the camera is allowed to have any inclination theoretically. (in fact, the user is recommended to use the inclination of less than 45 degrees or near 90 degrees). The time is 15 minutes or more, and the average accuracy is >93%.
Excellent ability to adapt to the environment
Module is not sensitive to the change of light and shadow caused by stream of people, season and climate. Therefore, it can not only adapt to indoor projects, but also to outdoor.
In the current version, we can detect the presence of umbrellas in the outdoor area to prevent summer omissions, which makes it possible to maintain 85% accuracy at the time of 15 minutes or more.
Strengthen learning ability
There is still a clear gap between the number of statistics system and the ability of human recognition, so the difference in application scenarios has an impact on the performance of the number of statistical systems. Like human vision system, our number counting device also has acquired learning ability. If we add new samples of the local area to retrain the detector, the performance of the algorithm will further improve.