Validation study demonstrates superior behavior detection in classroom video datasets, indicating strong potential for scalable intelligent education analytics.
Key Points
To develop and validate a specialized deep neural network, FGD-Net, designed to accurately detect fine-grained individual student behaviors despite occlusion, subtle visual cues, and classroom clutter.
Engineered an architecture featuring dynamic recalibration convolution blocks for local postures, multi-path gated context aggregation for contextual cues, and shift-guided upsampling for small spatial details.
Evaluated performance benchmarks against existing state-of-the-art object and action detectors on the SCB5 and SCB3 classroom behavior detection datasets.
Achieved superior individual behavior detection performance compared to recent modern detector baselines across both the SCB5 and SCB3 evaluation benchmarks.
Demonstrated qualitative improvements in capturing small-scale behavior cues and distinguishing visually similar upper-body gestures in dense classroom environments.