Recently, we are working on Face Tracking Based on meanshift. The effect is average. The hue component is selected as the feature in the standard algorithm. To improve the robustness of the background, a multi-feature space is proposed based on features such as gradients and bps. However, the histogram dimension is too small and spatial information is lost, which seriously degrades the value of feature classification. After testing, the trace becomes invalid when the background color is similar to the skin color (yellow. So I looked at how the products on the market made traffic statistics. Human traffic monitoring generally uses the camera ceiling installation method, in order to avoid problems such as congestion and overlap when the traffic volume is large.
References:
1: http://www.eccn.com/design_2011101810512969.htm. An embedded solution that uses hog features for head detection. It is divided into two parts: Training and detection. SVM is used as the classifier. Similar to the wen'an solution, wen'an's demo video is ineffective. Question and thinking: how to deal with the training classifier wearing a hat? If the background on the ground is similar to the main color of the human head (relatively small ). 2. Use the characteristics of the combination of the header and shoulder (similar to the APRIs solution, the APRIs demo is good, of course, the environment is relatively simple ). Http://cdmd.cnki.com.cn/Article/CDMD-10255-1012312394.htm Haar and hog classifier, from coarse to fine, improve the speed. Classifier features: hog (using structural symmetry) Training Mode: Adaboost, classifier SVM cascade 3, source code http://www.pudn.com/downloads456/sourcecode/graph/texture_mapping/detail1919552.html seems to have trained classifier, tracking algorithm for particle filtering. Http://www.doc88.com/p-217659743554.html problems and thinking: Some Solutions use motion detection, only deal with the part of motion. Does it affect head stop (pause? Where is the Classifier Training sample from? Is training unified, or are different samples collected for training based on different installation scenarios?
Http://blog.sina.com.cn/s/blog_49d1bc360101243t.html
Pattern Recognition Development Project-human traffic monitoring based on Head Detection