• Laser & Optoelectronics Progress
  • Vol. 57, Issue 22, 221016 (2020)
Hansong Su, Tengteng Liu, Gaohua Liu*, and Xichu Tian
Author Affiliations
  • School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
  • show less
    DOI: 10.3788/LOP57.221016 Cite this Article Set citation alerts
    Hansong Su, Tengteng Liu, Gaohua Liu, Xichu Tian. Algorithm for Student Behavior Detection Based on Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(22): 221016 Copy Citation Text show less
    Improved O-Net network
    Fig. 1. Improved O-Net network
    Overall network structure
    Fig. 2. Overall network structure
    Activation functional curves
    Fig. 3. Activation functional curves
    Loss functional curves
    Fig. 4. Loss functional curves
    Dataset examples. (a)(b)Stand; (c)(d) hand up; (e)(f) listen; (g)(h) write; (i)(j) sleep; (k)(l) multiple students
    Fig. 5. Dataset examples. (a)(b)Stand; (c)(d) hand up; (e)(f) listen; (g)(h) write; (i)(j) sleep; (k)(l) multiple students
    Flow chart of behavior detection on Jetson TX2
    Fig. 6. Flow chart of behavior detection on Jetson TX2
    Detection accuracy of five kinds of behaviors
    Fig. 7. Detection accuracy of five kinds of behaviors
    LayerOutput sizeKsizePaddingStride
    Conv 1100×1005×5Same1
    Conv 250×502×2Valid2
    Conv 350×505×5Same1
    Max pool25×252×2Valid2
    Conv 425×253×3Same1
    Conv 512×123×3Valid2
    Conv 612×123×3Same1
    Conv 76×62×2Valid2
    Fc 11152
    Fc 2192
    Fc 35
    Table 1. Network structure for image classification
    Network structurePreluh-Swishbh-Swish
    SEUCLIDRelative steep lossSEUCLIDRelative steep lossSEUCLIDRelative steep loss
    Original MTCNN75.375.977.677.878.278.3
    Improved MTCNN75.876.277.878.178.478.6
    Table 2. Accuracies of different network structures under different activation functions and loss functions unit: %
    AlgorithmYOLOv3tiny-YOLOv3SSDMobileNetV2-SSDMobileNetV3-SSDOur algorithm
    Dataset78.266.779.275.877.278.6
    Jetson TX274.560.477.173.374.177.9
    Table 3. Accuracy of each algorithm unit: %
    AlgorithmYOLOv3tiny-YOLOv3SSDMobileNetV2-SSDMobileNetV3-SSDOur algorithm
    Time78088653807669
    Table 4. Average time of each algorithm to process 1 frame image on Jetson TX2 unit: ms
    Hansong Su, Tengteng Liu, Gaohua Liu, Xichu Tian. Algorithm for Student Behavior Detection Based on Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(22): 221016
    Download Citation