Classroom video frequency point arrival method combining face recognition technology and pedestrian recognition technology
Technical Field
The invention relates to the field of face recognition, in particular to a classroom video frequency point arrival method combining face recognition and pedestrian recognition technologies.
Background
In a traditional classroom, teachers and educational administration department want to quickly and accurately know which students are not in class, so that subsequent teaching management is implemented. If the teacher calls one student, much time is wasted, and in the case of 50 students, the time is about 3-5 minutes, which is not always accurate. If the student signs his or her own signature, the student signs his or her own signature. If the attendance machine with a card swiping function, a fingerprint function or a human face is arranged, students need to queue to check in before or after class, which is very troublesome and additionally increases the cost.
A face recognition point arrival method based on video monitoring is already available in the market, and on the representative of monitoring manufacturers such as Haikangwei video and the like, the feature extraction is carried out on the face captured in real time by a monitoring picture in the background, and a face library of a student list is carried out by the steps of 1: n, searching, confirming one by one, and finally realizing no perception point. The background server decodes the video stream of the monitoring camera and then performs face recognition on the image frame by frame or frame by frame, namely a dynamic recognition method, but the method depends heavily on whether the student is looking ahead or not, and especially when the student always lowers the head or the face angle is more inclined, the situation of missed grabbing occurs. In a classroom with about 100 persons, the result is output within 5 minutes, the recognition rate of students is about 80%, and if the time is widened to 10 minutes, the recognition rate is increased to 85-90%.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a classroom video frequency point arrival method combining face recognition and pedestrian recognition technologies, and the specific technical scheme is as follows:
a classroom video frequency arrival method combining face recognition and pedestrian recognition technologies comprises the following steps: the following steps are adopted for the preparation of the anti-cancer medicine,
step 1: the processing module carries out pedestrian re-identification detection, people counting, modeling classroom position distribution map and position coordinates of each person on a single picture at the time t in the video stream of the camera; setting the set of students in class as N, the number of the detected pedestrians is N ═ N |, the set of the pedestrians is identified as M (t), and the number of the pedestrians is m (t), wherein m (t) is less than or equal to N;
step 2, extracting the features of the human face in the picture, retrieving a class list according to the human face features, confirming a current set K (t), iterating with a previous picture human face confirmation set K (t-1) to obtain a confirmation set K (t) K (t) ∪ K (t-1) at the time t, sequentially identifying the human features of each person in the set K (t) by a human feature identification module, wherein the human features comprise hairstyle, clothing, wearing features and height, establishing a corresponding relation between the identified human features and the human face features by a processing module, and storing the identified human features in a human feature database;
and step 3: judging whether the number of people in K (t) is equal to n, if so, entering a step 8, otherwise, entering a step 4;
and 4, step 4: if the number of people in K (t) is less than n, sequentially carrying out human feature probability estimation on each member subject is in the set M (t) -K (t);
the prior attendance probability P (ia), the is ∈ M (t) -K (t) and the ia ∈ O of the member objects ia in the set O are obtained from a database in advance by setting the set O to be N-K (t);
according to the identified hair style of the target is, searching the association characteristics in a human body characteristic database aiming at the member object ia to obtain a similarity probability P1 (is);
according to the identified clothing of the target is, searching the associated features in a human body feature database aiming at the member object ia to obtain a similarity probability P2 (is);
according to the identified wearing feature of the target is, searching a human body feature database aiming at the member object ia to obtain a similarity probability P3 (is);
sequentially taking an estimation probability set Q ({ Ps) ((ia) }) P1(is) } P2(is) } P3(is) | P4(is) | ia ∈ O, is } for each member object is in M (t) -K (t), and selecting a corresponding estimation attendee with the maximum value in the estimation probability set Q as the member object is to obtain a set B (t) consisting of the estimation attendees;
and 5: setting iteration times for at least p times, re-analyzing a new picture at the moment of t +1, and repeating the steps 1 to 4 for at least p times;
step 6: after iterating the set p pictures, if the termination condition is not reached, stopping iteration, and entering step 7, otherwise, entering step 8;
and 7: updating the prior attendance probability and the human body feature library probability of each member, and outputting M (t), K (t) and B (t), wherein M (t) is K (t) + B (t);
and 8: the confirmation of all students is completed, the round of checking is finished, and the output result is the student attendance list N (t), K (t), M (t).
Further: in the step 1, a deep learning-based CNN network is adopted to identify a single picture at the time t in a video stream of a camera.
The invention has the beneficial effects that: firstly, hardware equipment and a camera do not need to be modified, most of the existing cameras can meet the requirements, and the user cost is reduced;
secondly, the accuracy and speed of the original face recognition dynamic point arrival are improved. The invention can solve the problem of fast and efficient arrival of the students in the classroom at present.
Drawings
FIG. 1 is a flow chart of the operation of the present invention.
Detailed Description
The following detailed description of the preferred embodiments of the present invention, taken in conjunction with the accompanying drawings, will make the advantages and features of the invention easier to understand by those skilled in the art, and thus will clearly and clearly define the scope of the invention.
As shown in fig. 1: a method for detecting classroom video frequency by combining face recognition and pedestrian recognition technology comprises the following steps,
step 1: the processing module carries out pedestrian re-identification detection, people counting, modeling classroom position distribution map and position coordinates of each person on a single picture at the time t in the video stream of the camera;
setting the set of students in class as N, the number of the detected pedestrians is N ═ N |, the set of the pedestrians is identified as M (t), and the number of the pedestrians is m (t), wherein m (t) is less than or equal to N;
step 2, extracting the features of the human face in the picture, retrieving a class list according to the human face features, confirming a current set K (t), iterating with a previous picture human face confirmation set K (t-1) to obtain a confirmation set K (t) K (t) ∪ K (t-1) at the time t, sequentially identifying the human features of each person in the set K (t) by a human feature identification module, wherein the human features comprise hairstyle, clothing, wearing features and height, establishing a corresponding relation between the identified human features and the human face features by a processing module, and storing the identified human features in a human feature database;
and step 3: judging whether the number of people in K (t) is equal to n, if so, entering a step 8, otherwise, entering a step 4;
and 4, step 4: if the number of people in K (t) is less than n, sequentially carrying out human feature probability estimation on each member object i s in the M (t) -K (t) set;
the prior attendance probabilities P (ia), i s ∈ M (t) -K (t), and i a ∈ O of the member objects i a in the set O are obtained from the database in advance by setting the set O to N-K (t);
according to the identified hair style of the target i s, searching the association characteristics in a human body characteristic database aiming at the member object ia to obtain a similarity probability P1(i s);
according to the clothing of the identified target i s, searching the human body characteristic database for the associated characteristics aiming at the member object ia to obtain a similarity probability P2(i s);
according to the recognized wearing characteristics of the target i s, searching a human body characteristic database aiming at the member object ia to obtain a similarity probability P3(i s);
according to the identified height characteristics of the target i s, searching the human body characteristic database aiming at the member object ia to obtain a similarity probability P4(i s);
for each member object i s in M (t) -K (t) in turn,
taking an estimation probability set Q ═ { P s ═ P (ia) ═ P1(i s) × P2(i s) × P3(i s) × P4(i s) | i a ∈ O, i s }, selecting the maximum value in the estimation probability set Q as the corresponding estimation attendee of the member object i s, and obtaining a set b (t) consisting of estimation attendees;
and 5: setting iteration times for at least p times, re-analyzing a new picture at the moment of t +1, and repeating the steps 1 to 4 for at least p times;
step 6: after iterating the set p pictures, if the termination condition is not reached, stopping iteration, and entering step 7, otherwise, entering step 8;
and 7: updating the prior attendance probability and the human body feature library probability of each member, and outputting M (t), K (t) and B (t), wherein M (t) is K (t) + B (t);
and 8: the confirmation of all students is completed, the round of checking is finished, and the output result is the student attendance list N (t), K (t), M (t).