Figures (7)  Tables (6)

    • Figure 1. 

      A drone recording the motion of vehicles along a 420-meter stretch of highway from an overhead perspective[38].

    • Figure 2. 

      The illustration of spatial range from a top view.

    • Figure 3. 

      The detailed and simplified top views for each moment with the subject vehicle in the center. (a) is an illustration of the spatial range, (b) is a zoomed-in version of the study area, and (c) is a simplified version as input to the model.

    • Figure 4. 

      The time series top view set describing an event.

    • Figure 5. 

      The structure of CNN model.

    • Figure 6. 

      The basic unit of LSTM model[43].

    • Figure 7. 

      The proposed CNN-LSTM model.

    • Modeling processParameters with values
      Input of CNN75 top views with both the front and rear of the subject vehicle: 360 × 30
      (or 75 top views with only the front of the subject vehicle: 360 × 60)
      Convolution layerNo. of layers: 4
      No. of kernels: 32, 64,128, and 256
      Kernel size: (5 × 5), (3 × 3), (3 × 3), and (3 × 3)
      Stride: (2,2), (2,2), (2,2), and (2,2)
      Padding: (0,0), (0,0), (0,0), and (0,0)
      Activating function: ReLU
      Pooling layerNo. of layers: 1
      Kernel size: (2 × 2)
      Stride: (2,2)
      Padding: (0,0)
      Fully connected layerNo. of layers: 2
      Hidden neurons: 512 and 256
      Activating function: ReLU
      Output of CNN model/ Input of LSTM modelNo. of features: 64 (for each top view)
      75 top views for each event
      LSTMNo. of layers: 3
      Hidden neurons: 512, 512, 512
      Fully connected layerNo. of layers: 1
      Hidden neurons: 256
      Activating function: ReLU
      Output of LSTM modelBinary classification result: high-risk or non-high-risk
      Training processBackpropagation
      Learning rate: StepLR (lr = 1e-3, γ = 0.3)
      Loss function: Cross-entropy
      Mini-batch size: 128
      Epochs: 50

      Table 1. 

      Parameters with values in the CNN-LSTM modeling process.

    • Actual conditionPredicted condition
      PositiveNegative
      PositiveTrue Positive (TP)False Negative (FN)
      NegativeFalse Positive (FP)True Negative (TN)

      Table 2. 

      The confusion matrix.

    • Spatial rangeTemporal range
      1100 m in both the front and rear of the subject vehicle5 s to 2 s before the zero time
      25 s to 3 s before the zero time
      34 s to 2 s before the zero time
      4Only 100 m in the front of the subject vehicle5 s to 2 s before the zero time
      55 s to 3 s before the zero time
      64 s to 2 s before the zero time

      Table 3. 

      Descriptions of the six experiment designs.

    • SensitivityFalse Alarm RateAUC
      10.9880.1770.992
      20.9770.1190.988
      30.9880.2740.989
      40.9960.0650.997
      50.9920.0320.996
      60.9880.3870.988

      Table 4. 

      The prediction performances of the six experiment designs.

    • Loss and accuracyROC curve
      1
      2
      3
      4
      5
      6

      Table 5. 

      The loss, accuracy, and ROC curve of the six experimental designs.

    • ModelVariable dimensionTime-series variability considerationSensitivityFalse alarm rateAUC
      Random Forest[44]Cross-sectionNo0.5170.0880.827
      RPLR model[40]Single vehicleYes0.9800.0600.960
      CNN-LSTM model (this study)Single vehicleYes0.9960.0650.997

      Table 6. 

      Comparation results of modeling performance based on testing data.