Design And Implementation Of Head Detection System Under Complex Environment

With the development and progress of modern science, artificial intelligence has been widely used in various fields because its excellent performance in terms of saving resources, improve efficiency, safety and quality assurance, etc.. Machine vision is a major research focus in artificial intelligence, in particular, plays a key role in security aspects.Techniques of pedestrian detection and tracking, face detection, expression analysis, pose estimation are becoming increasingly sophisticated currently, various detection algorithms have been devised. Because the human body in complex scenes is susceptible to light, texture, and varying degrees of occlusion, deformation, relatively speaking, while the head contains a very rich feature information, and rarely long obscured or deformed, so fast head detection provides a practical research to achieve a real-time target detection, tracking, counting.Head detection is to identify and locate the head in static or sequence of images, on the basis of domestic and foreign technology development and research literature we proposed two kinds of head detection methods based on statistical learning, which are detection methods with feature classification and convolution neural network.Target detection method has been developed based on the feature classification relatively mature, but people often need to do a good tradeoff between the detection accuracy and detection rate. In feature-based detection of the proposed classification, we select the NR-LBP and HOG features to build a cascade classifier and make double image detection to meet the needs of real-time and ensure the detection accuracy of the premise, which has been proved to have good robustness with illumination and has a high detection speed.In the detection method based on convolution neural network, we design a seven-layer neural network to do the feature extraction and classification. The structure of convolution neural network concludes close-connecting layers and space information which make it easy applied especially to the image processing and pattern recognition. Experiments show that the method is more superior with the picture rotation, translation, deformation than the other algorithms, and the test results are accurate and reliable.This article better ensures the accuracy and real-time synchronization by providing two ideas of the head detection method based on statistical learning, and can basically meet the market requirements for head detection.