Applied Science and Innovative Research 

ISSN 2474-4972 (Print) ISSN 2474-4980 (Online) 

Vol. 8, No. 2, 2024 

www.scholink.org/ojs/index.php/asir 

93 
 

Original Paper 

Student Classroom Behavior Detection Algorithm Based on 

Improved YOLOv8n 

Huiming Li 

School of Applied Technology, University of Science and Technology Liaoning, Anshan, 114051, 

China 

 

Received: May 5, 2024           Accepted: May 25, 2024          Online Published: May 29, 2024 

doi:10.22158/asir.v8n2p93           URL: http://doi.org/10.22158/asir.v8n2p93 

 

Abstract 

Currently, object detection technology is becoming more and more mature, but there are still a lot of 

challenges in recognizing students' classroom behaviors. In order to address the problems of 

inaccuracy and high computation of existing models in the process of student classroom behavior 

recognition, this paper adopts the improved Yolov8n model to detect and recognize student classroom 

behavior. Based on Yolov8n, the method adds an efficient shuffle attention to increase the ability of 

feature extraction and improve the model recognition accuracy; secondly, the bounding box loss 

function is optimized to improve the model's localization ability. The experimental results show that the 

mAP50 and mAP50-95 metrics of the proposed model on the student classroom behavior dataset are 

99.9% and 95.4%, respectively. The proposed model can achieve the detection and identification of 

classroom behavior more quickly and accurately with lower computing cost, and can realize the 

dynamic and scientific identification of students' classroom learning. 

Keywords 

student classroom behavior, YOLOv8, object detection, CNNs  

 

1. Introduction 

Students' classroom performance is an important part of classroom teaching evaluation, and the 

identification of students' classroom behavior is of great significance for classroom teaching reform and 

exploration. The classroom behavior detection method based on artificial intelligence and deep learning 

can effectively identify the typical classroom behaviors of students and help teachers to grasp the 

learning status of students in a timely and effective manner, thus contributing to intelligent classroom 

teaching. 

Traditional classroom teaching behavior analysis mainly consists of two parts: feature extraction and 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

94 
Published by SCHOLINK INC. 

classification, such methods require human judgment in the recognition process and are affected by 

other factors such as the detection target and specific tasks, resulting in low detection accuracy, and the 

detection of complex backgrounds has certain limitations, which makes it difficult to promote and 

expand on a large scale. In recent years, with the development of deep learning, convolutional neural 

networks (CNNs) have become the mainstream of detection and recognition technology with excellent 

feature extraction capabilities. Currently, mainstream object detection and recognition algorithms are 

divided into two major categories: a two-stage approach based on region suggestion and a one-stage 

approach based on object regression. R-CNN (Girshick et al., 2014) is a two-stage object detection 

algorithm, which uses a region suggestion network to extract candidate regions from an image, and 

then recalibrates the objects in the candidate regions to obtain the detection results. Subsequently, 

improved two-stage algorithms such as Fast R-CNN (Girshick, 2015), Faster R-CNN (Ren et al., 2016), 

and Mask R-CNN (Gkioxari, 2017) have been proposed to improve detection accuracy and speed. 

However, two-stage target detection algorithms still have the problem of slow detection speed. 

One-stage object detection algorithms are significantly faster than two-stage networks because do not 

need region suggestion and get the detection results directly by end-to-end, typical algorithms are SSD 

(Liu et al., 2016) and YOLO (Redmon et al., 2016; Redmon et al., 2017) series algorithms. YOLO 

series detection algorithms have gone through several version updates, which ensure the detection 

speed of the algorithms while improving the detection accuracy and therefore widely used in the field 

of object detection. 

The recent research of student classroom behavior detection has achieved more accurate detection goals, 

but there are still problems such as small sample size, weak generalization ability, and slow inference 

speed. In order to solve these problems, this paper is based on the newly released YOLOv8n model in 

YOLO series, which has higher accuracy and smaller number of parameters, and improves and 

optimizes the method of students' classroom behavior detection, in order to develop a detection model 

with high accuracy and fast detection speed for resource-constrained devices. 

 

2. Materials and Methods 

2.1 Dataset 

The data in this paper is from Student Behavior Detection Dataset (Burak, 2024), we have filtered and 

uniformly sampled the dataset, and collated a total of 2626 pieces of student classroom behavior image, 

the resolution of each image is 640×640 pixels, and the number of various behavior categories are: 

forward, 459; hand, 495; read, 403; sleep, 423; around, 482. The annotation format is YOLO format. 

The dataset is divided into training set, validation set and test set in the ratio of 7:2:1. Typical student 

classroom behaviors are shown in Figure 1. 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

95 
Published by SCHOLINK INC. 

 

Figure 1. Typical Student Classroom Behaviors 

 

2.2 YOLOv8 Architecture 

In the field of object detection, YOLO series models are widely used because of their high accuracy 

and small model size. YOLOv8 consists of three parts, Backbone, Neck and Head, and its architecture 

is shown in Figure. 2. Backbone is responsible for extracting high-level semantic features from the 

input images. Neck is located between Backbone and Head, and is used to perform multi-level feature 

fusion to enhance the network's ability to perceive the target. The detection head is responsible for 

localizing and classifying the target to generate the final detection result. Compared with YOLOv5, 

YOLOv8 replaces the C3 module with the C2f module, which is lighter and makes the model more 

adaptable to different sizes and shapes of targets, and adopts the current mainstream decoupled head 

structure, which effectively reduces the number of parameters and computational complexity, and 

enhances the model's generalization ability and robustness. In addition, YOLOv8 abandons the design 

of using Anchor-Base to predict the position and size of the anchor box in the previous series, and uses 

Anchor-Free to directly predict the center point and width-to-height ratio of the target, reducing the 

number of anchor boxes, which further improves the detection speed and accuracy of the model.  



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

96 
Published by SCHOLINK INC. 

 

Figure 2. YOLOv8 Architecture 

 

2.3 Network Model Improvement 

2.3.1 Add Shuffle Attention 

There should be significant differences in features between different targets or between targets and 

background, which is crucial for effectively identifying and locating targets. The object detection 

algorithm will recognize these features in a certain context and find regions in the image that match 

these features. Although the YOLOv8 model can extract detailed behavioral features of students in 

classroom scenarios, but classroom scenarios have complexity and variability, so how to be able to 

learn important features in the scenarios has become a problem that must be solved. Shuffle Attention 

(SA) (Zhang et al., 2021) uses replacement units to efficiently combine channel and spatial attention, to 

solve the problem of the high computational complexity of the combined attention. The architecture of 

SA is shown in Figure 3. 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

97 
Published by SCHOLINK INC. 

 

Figure 3. The Architecture of SA 

 

Firstly, the input feature 𝑋 ∈ 𝑅(𝐶×𝐻×𝑊)  divided into G groups along the channel dimension, 

represented as 

𝑋 = [𝑋1, 𝑋2, … , 𝑋𝐺], 𝑋𝑘 ∈ 𝐶 𝐺⁄ × 𝐻 ×𝑊                       (1)  

Each sub feature Xk captures a specific semantic information. Next, Xk is divided into two branches 

along the channel dimension, with two sub features represented as 𝑋𝑘1, 𝑋𝑘2 ∈ 𝑅(𝐶 2⁄ 𝐺×𝐻×𝑊), Xk1 

branch captures inter channel dependencies, The Xk2 branch captures spatial dependencies, and 

achieves attention to semantic and positional information through two branches. Subsequently, the two 

attentions are concatenated to obtain features with the same input dimension. The above operation can 

be achieved through the following formula: 

𝑠 = 𝐹𝑔𝑝(𝑋𝑘1) =
1

𝐻×𝐾
∑ ∑ 𝑋𝑘1(𝑖, 𝑗)

𝑊
𝑗=1

𝐻
𝑖=1                         (2) 

𝑋𝑘1
′ = 𝜎(𝐹𝑐(𝑠)) ∙ 𝑋𝑘1 = 𝜎(𝑊1𝑠 + 𝑏1) ∙ 𝑋𝑘1                       (3) 

𝑋𝑘2
′ = 𝜎(𝑊2 ⋅ 𝐺𝑁(𝑋𝑘2) + 𝑏2) ⋅ 𝑋𝑘2                           (4) 

𝑋𝑘
′ = [𝑋𝑘1

′ , 𝑋𝑘2
′ ] ∈ 𝑅𝐶 𝐺⁄ ×𝐻×𝑊                            (5) 

where, σ represents the activation function, GN represents group normalization. Finally, an attention 

map with the same dimension as the input is obtained through feature aggregation operation. 

The main task of YOLOv8's backbone network is to extract features. Therefore, we add SA before the 

SPFF module of the backbone network to enhance the network's ability to autonomously learn key 

features and discard other unimportant features. 

2.3.2 Optimizing the Bounding Box Loss Function 

The bounding box loss function, as an important component of the object detection loss function, is 

crucial for the detection performance of the model. YOLOv8 uses CIoU to calculate the regression loss 

of bounding boxes, but it has issues such as high parameter sensitivity, computational complexity, and 

slow convergence speed. In order to overcome the above problems, this paper replaces the CIoU in the 

baseline model with WIoU (Tong et al., 2023) based on dynamic non monotonic focusing mechanism 

to improve the prediction speed and accuracy of the model. The calculation process is shown in 

equation (6): 

𝐿𝑊𝐼𝑜𝑈 = 𝑟𝑅𝑊𝐼𝑜𝑈𝐿𝐼𝑜𝑢, 𝐿𝐼𝑜𝑢 ∈ [0,1], 𝑅𝑊𝐼𝑜𝑈 ∈ [1, 𝑒]                    (6) 

where, LIoU represents IoU loss, which will weaken the penalty term of high-quality anchor boxes and 

strengthen its focus on the centroid distance when the overlap between the anchor box and the 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

98 
Published by SCHOLINK INC. 

predicted box is high; RWIoU represents the penalty term of WIoU, used to strengthen the loss of 

ordinary quality anchor boxes; The superscript * represents not participating in backpropagation, 

effectively preventing the network from generating gradients that cannot converge; r dynamically 

allocate the gradient gain of bounding boxes and reduce harmful gradients generated by low-quality 

anchor boxes in the later stages of training, focusing on ordinary quality anchor boxes to improve the 

model's localization ability. RWIoU and r are defined as shown in equations (7)-(8). 

𝑅𝑊𝐼𝑜𝑈 = 𝑒𝑥𝑝 (
(𝑥−𝑥𝑔)

2
+(𝑦−𝑦𝑔)

2

(𝑊𝑔
2+𝐻𝑔

2)
∗ )                          (7) 

𝑟 =
𝛽

𝛿𝛼𝛽−𝛿
                                   (8) 

where the parameters (x, y）, (xg, yg）, (Wg, Hg) represents the center position of the predicted box, the 

center position of the ground-truth box, and the width and height of the minimum rectangle containing 

the predicted box and the ground-truth box, respectively. β is the outlier with a small gradient gain, 

which can reduce the impact on bounding box regression. α, δ denote hyperparameters. 

 

3. Experiment 

3.1 Experimental Environment and Datasets 

This experiment was conducted under the 64-bit operating system Ubuntu 23.10, the GPU was 

NVIDIA RTX 3060 with 12 GB of video memory, the host memory was 16 GB, the programming 

language was Python 3.11.6, the GPU was accelerated using CUDA v12.2, and the training was based 

on the deep learning framework Pytorch 2.1.0. The SGD optimizer was used, the initial learning rate 

was set to 0.01, the batch size was 32, and the input image size was 640×640. 

3.2 Metrics 

This experiment uses precision, recall, and mean average precision (mAP) to evaluate the model. The 

mAP50 is the average of the precision of all the samples with threshold 0.5, which reflects the trend of 

the model's precision with the recall rate, and the higher the value means that the model is more likely 

to maintain a high precision with a high recall rate. The mAP50-95 represents the average precision at 

different IoU thresholds (from 0.5 to 0.95 in 0.05 steps). 

3.3 Experimental Results and Analysis 

Figure 4 shows the loss curve and validation performance of the model during training and validation. 

From the graph, it can be seen that as the number of training iterations increases, the training and 

validation losses of the model gradually decrease, indicating that the model continuously learns more 

accurate features. From the perspective of validation effectiveness, the model performs well in terms of 

precision rate, recall rate, mAP50 and mAP50-95. Figure 5 shows the prediction results of the model 

on the test set. It can be seen from the figure that the model correctly predicted the classroom behavior 

of students with high confidence. 

 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

99 
Published by SCHOLINK INC. 

 

Figure 4. Loss, Accuracy, and Recall of the Model in the Training and Validation Sets 

 

 

Figure 5. Prediction Results of the Model on the Test Set 

 

3.4 Ablation Experiment 

To verify the optimization effect of each module, ablation experiments were conducted in this paper, 

and the results are shown in Table 1. According to Table 1, after adding SA to the baseline model, 

MAP50-95 increased by 0.06. After introducing the regression loss function WIoU, MAP50-95 

increased by 0.02. The proposed model shows higher accuracy in detecting student classroom behavior 

compared to the YOLOv8n baseline model, proving the effectiveness and feasibility of the optimization 

module. 

 

 

 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

100 
Published by SCHOLINK INC. 

Table 1. The Ablation Experiment Results 

Model precision recall mAP50 mAP50-95 

Baseline 0.999 1 0.995 0.946 

Baseline+SA 0.999 1 0.995 0.952 

Baseline+SA+ WIoU 0.999 1 0.995 0.954 

 

4. Conclusion 

Students' behavior in the classroom reflects the effect of classroom teaching, and also reflects their 

participation in the course, and the research and analysis of classroom behavior can provide direction 

for the evaluation of classroom teaching. In this paper, we propose a classroom behavior detection 

algorithm based on the improved YOLOv8n model, which improves the feature learning ability of the 

model by adding the attention mechanism to the baseline model, and improves the accuracy and speed 

of the model by introducing the WIoU bounding box loss function. The experimental results show that 

the proposed algorithm can effectively identify and locate students' classroom behaviors, which has 

certain teaching aid value. 

 

References 

Burak. (2024). Student Behavior Detection Dataset. 

https://universe.roboflow.com/burak-koyfx/student-behavior-detection 

Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE international conference on computer 

vision (pp. 1440-1448). https://doi.org/10.1109/ICCV.2015.169 

Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object 

detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision 

and pattern recognition (pp. 580-587). https://doi.org/10.1109/CVPR.2014.81 

He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask r-cnn. In Proceedings of the IEEE 

international conference on computer vision (pp. 2961-2969). 

https://doi.org/10.1109/ICCV.2017.322 

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C. Y., & Berg, A. C. (2016). Ssd: Single 

shot multibox detector. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, 

The Netherlands, October 11–14, 2016, Proceedings, Part I 14 (pp. 21-37). Springer International 

Publishing. https://doi.org/10.1007/978-3-319-46448-0_2 

Redmon, J., & Farhadi, A. (2018). Yolov3: An incremental improvement. arxiv preprint 

arxiv:1804.02767. 

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time 

object detection. In Proceedings of the IEEE conference on computer vision and pattern 

recognition (pp. 779-788). https://doi.org/10.1109/CVPR.2016.91 



www.scholink.org/ojs/index.php/asir             Applied Science and Innovative Research                  Vol. 8, No. 2, 2024 

101 
Published by SCHOLINK INC. 

Ren, S., He, K., Girshick, R., & Sun, J. (2016). Faster R-CNN: Towards real-time object detection with 

region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6), 

1137-1149. https://doi.org/10.1109/TPAMI.2016.2577031 

Tong, Z., Chen, Y., Xu, Z., & Yu, R. (2023). Wise-IoU: bounding box regression loss with dynamic 

focusing mechanism. arxiv preprint arxiv:2301.10051. 

Zhang, Q. L., & Yang, Y. B. (2021 June). Sa-net: Shuffle attention for deep convolutional neural 

networks. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal 

Processing (ICASSP) (pp. 2235-2239). IEEE. 

https://doi.org/10.1109/ICASSP39728.2021.9414568