Crowd Abnormal Behavior And Face Alignment Algorithm Design

With rapid development of society and economy, the extent to urbanization and technicalization is deeper and deeper. The large-scale crowd activities are more and more widespread. As a result, some abnormal crowd phenomenon such as tread, panic, riot and terrorist attack are quite frequent. Therefore, how to supervise and control large-scale crowd activities effectively has become the key point research in the area of city safety security.Detection of crowd abnormal behaviors mainly studies the crowd phenomenon and its movement creatures. Recently, computer vision, pattern recognition and machine learning technologies etc. have gradually developed, which makes the studies of crowd abnormal behaviors being applied to the intelligent supervisory equipment terminal.On the base of summering and analyzing former crowd abnormal recognition algorithms, my essay proposes and completes a set of crowd abnormal behaviors recognition and crowd density estimate algorithms, which can be applied to the detection of different occasions such as abnormal movements, panic and scenes that require crowd density estimate.As the aspect of abnormal movement detection, the main purpose of which is to detect individuals who are running or retrograding in normal crowd. The process is to compute the moving velocity and direction of each movement unit by combining dense optical flow and interpolation method. And compare it with the moving direction and amount of the whole picture. If the difference is quiet large, it can be marked as abnormal movement. Referring to the part of panic detection, the main purpose is to detect the crowd abnormal behaviors such as riot, panic etc. The general process is that, first, compute the frame difference between former and later frames, and get changed frame points. Then, calculate the second-order differential of each changing frame points, and get the spatial feature points. Later, calculate 3*3 neighborhood's average displacement of each feature point and count the changes between two continuous frames. Finally, study the samples of normal and abnormal situations by using support vector machine(SVM) to recognize abnormal crowd behaviors. The algorithm can also be applied to achieve the recognition of crowd density.On the base of the design of crowd abnormal detection algorithm, my internship also studies facial recognition algorithm.Studying of facial recognition is to perfect the relevant functions of crowd abnormal recognition. As for crowd or individual videos caught by the video camera, facial recognition can be achieved by relevant algorithms so as to conduct statistics of persons appear on the shot. On the other side, by setting white list or black list in cameras of some pivotal fields, the passers-by can be restricted in specific areas.Face alignment discussed in my essay is based on facial area has been recognized successfully and intercepted by other algorithms designed by my colleges. However, the key points in intercepted facial area such as eyes, nose, angulus oris etc. require being located. The located facial picture and its key point's data can be applied to the later facial recognition algorithm. The Face Alignment process is achieved by using Deep Multi-task Learning.My essay will introduce the experiments contrast of above algorithms in difference data sets in detail in order to prove its robustness and instantaneity.