Research on Intelligent Recognition Algorithm of College Students' Classroom Behavior Based on Improved SSD

This paper takes the classroom images of more than 50 classrooms in a university for nearly 4 months in a semester as the research object. The LabelImg manual annotation method is used to construct a detection data set including four behavioral states: listening to class, taking notes, playing with...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2022 IEEE 2nd International Conference on Computer Communication and Artificial Intelligence (CCAI) s. 160 - 164
Hlavní autoři: Wenchao, Lv, Meng, Huang, Yuping, Zhang, Shuai, Liu
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 06.05.2022
Témata:
Annotations
classroom behavior recognition
Data models
deep learning
Learning (artificial intelligence)
Manuals
Object detection
Surveillance
Target recognition
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:This paper takes the classroom images of more than 50 classrooms in a university for nearly 4 months in a semester as the research object. The LabelImg manual annotation method is used to construct a detection data set including four behavioral states: listening to class, taking notes, playing with mobile phones, and sleeping. In order to effectively improve the detection accuracy of the data annotation model, we used data enhancement techniques such as cropping, rotation, and shading transformation to expand the number of dataset. Based on this dataset, an improved SSD model based on deep learning target detection technology is adopted. ResNet module is used to solve the problem that VGG module has poor detection results of students' behavior state in picture analysis, FPN module is added to build RF-SSD detection model to improve the efficiency of image recognition to solve the problem of the low efficiency of small target recognition in the back of class. The experimental results show that RF-SSD has a great improvement in feature extraction ability and small target recognition accuracy compared with native SSD in self-constructed dataset, and can provide technical support and new ideas and methods for teaching management in universities.
DOI:10.1109/CCAI55564.2022.9807756