Research And Implementation Of People Count System Based On Machine Learning

Human population in the world is increasing dramatically. This growth, as a result from urbanization worldwide, has indirectly made crowd phenomenon increasing. The value of crowd analysis in public security,public infrastructure,business application has been showing. Compared to traditional statistical methods, machine learning method has accurate data processing, efficient information fusion and requires much fewer human operators. Therefore, intelligent visual surveillance at area under observation is extensively studied by computer vision researchers, as well as the emergence of a large number of methods to achieve people.Through the study of people count methods, the method based on statistical regression have a more accurate results, but count the number of people moving in different directions is more difficult to achieve. The method based on individual detection has obvious advantages on direction counts. However, due to its large computation in detection, the implementation on embedded system makes it impossible on real-time applications. Therefore, in this paper, based on the individual detection method, we proposed a method of target pre-fetch which extract the windows that containing suspected target to avoid full image scanning. Reuse subsequent detector for further testing on these windows to achieve the purpose of pedestrian detection and count while increasing processing speed.In this paper, people count system uses extraction method based on BING features. Experiments show that, the extracted windows based on BING feature have a great reference value. On pedestrian detection, in order to reduce errors due to occlusion, we use vertical overhead camera to recognize pedestrians by detecting human head. In the design of human head classifier, we use AdaBoost training method based on Haar-like features. In order to improve the detection performance, we use an algorithm for constructing a cascade of classifier. On human head tracking, our system uses the tracking method based on the prediction of Kalman filter. Algorithm simulation and debugging achieve on PC platform by using POSIX multi-thread technology. On system implementation. We select cubieboard2 based on Allwinner A20 processor as hardware platform. Finally, the system is ported to the hardware platform.Experimental results show that our extraction of human head based on BING feature has a good extraction effect while avoiding a large amount of computation caused by full image scanning. We use multi-thread technology to make the most of hardware processor resource. Finally, in order to verify the effectiveness of the system, we test our system in a number of scenarios. Test results show that our embedded people count system based on machine learning can accurately work in real-time application with overhead visual angle.