CBPH-Net: A Small Object Detector for Behavior Recognition in Classroom Scenarios

Abstract

Recognizing classroom behavior is crucial for assessing and improving teaching quality. However, existing methods for behavior recognition have limited accuracy due to issues such as occlusions, pose variations, and inconsistent target scales. To address these challenges, we propose an advanced single-stage object detector called CBPH-Net. Specifically, we design an efficient Feature Extraction Module (FEM) to capture more channel information and relevant features from the images in the backbone network. The neck network combines the PANet architecture and Coordinate Attention (CA) to integrate semantic and positional information and suppress irrelevant background information, enabling the network to accurately locate students. ConvNeXt Block Prediction Head (CBPH) utilizes convolutional kernels of different sizes and parsing multi-scale features to enhance the multi-scale recognition capability of CBPH-Net especially for accurate detection of small objects. To reduce the influence of irrelevant background, we use elliptical boxes instead of rectangular boxes when calculating the similarity between ground truth and predicted values. In addition, we construct a dataset named STBD-08 that contains 4432 images with 151574 labeled anchors covering 8 typical classroom behaviors. On the proposed dataset STBD-08, CBPH-Net achieves mean average precision (mAP) of 87.5% (an improvement of 3.4% compared to YOLOv5 and 1.2% compared to YOLOv7). It processes one frame with the latency of 31.3ms (1ms slower than YOLOv5 and 5.3ms faster than YOLOv7). Moreover, it achieves a precision of 75.7% in small object recognition, surpassing all comparative methods. The experimental results demonstrate that CBPH-Net can be efficiently applied to classroom behavior recognition tasks. Codes and datasets are available at: https://github.com/icedle/CBPH-Net.

Topics & Concepts