Scenario-based collision detection using machine learning for highly automated driving systems

Figures & data

Figure 1. Concept Model for using ML for HARA.

Figure 2. Representation of Use Case for Scene 1.

Figure 3. Parameter-based scenario modelling and optimization: (a) logical scenario: Parameter set; (b) logical scenario: Scenario Modelling.

Figure 4. Neural network with dense layers (exp_ a13 to exp_ a16).

Table 2. Summary of experimental results: deep neural network-MLP in HARA Use case.

	Summary of experiments
Model experiment${}^{\mathrm{a}}\mathrm{\_}\mathrm{I}\mathrm{D}$	Dense layers${}^{\mathrm{b}}$	K-fold	Input dataset${}^{\mathrm{c}}$	Avr. accuracy	Avr. loss
	layer.dense (8)
exp_a13	layer.dense (16)	5	dataset1	69.64%	0.1762
	layer.dense (8)
	layer.dense (8)
exp_a15	layer.dense (16)	10	dataset1	72.32%	0.1719
	layer.dense (8)
	layer.dense (8)
exp_a14	layer.dense (16)	5	dataset2	85.71%	0.1076
	layer.dense (8)
	layer.dense (8)
exp_a16	layer.dense (16)	10	dataset2	88.57%	0.0948
	layer.dense (8)
	layer.dense (16)
exp_a37	layer.dense (32)	5	dataset1	72.32%	0.1676
	layer.dense (8)
	layer.dense (16)
exp_a38	layer.dense (32)	10	dataset1	72.32%	0.1675
	layer.dense (8)
	layer.dense (16)
exp_a81	layer.dense (32)	5	dataset2	82.85%	0.1141
	layer.dense (8)
	layer.dense (16)
exp_a82	layer.dense (32)	10	dataset2	77.14%	0.1600
	layer.dense (8)
	layer.dense (32)
exp_a3	layer.dense (64)	5	dataset1	71.43%	0.1712
	layer.dense (32)
	layer.dense (32)
exp_a4	layer.dense (64)	10	dataset1	70.54%	0.1715
	layer.dense (32)
	layer.dense (32)
exp_a9	layer.dense (64)	5	dataset2	82.86%	0.1108
	layer.dense (32)
	layer.dense (32)
exp_a10	layer.dense (64)	10	dataset2	82.86%	0.1097
	layer.dense (32)
	layer.dense (32)
exp_a1	layer.dense (32)	5	dataset1	73.21%	0.1694
	layer.dense (32)
	layer.dense (32)
exp_a2	layer.dense (32)	10	dataset1	72.32%	0.1682
	layer.dense (32)
	layer.dense (32)
exp_a7	layer.dense (32)	5	dataset2	82.45%	0.1093
	layer.dense (32)
	layer.dense (32)
exp_a8	layer.dense (32)	10	dataset2	80.00%	0.1144
	layer.dense (32)
	layer.dense (32)
exp_a5	layer.dense (64)	5	dataset1	72.32%	0.1681
	layer.dense (128)
	layer.dense (32)
exp_a6	layer.dense (64)	10	dataset1	70.54%	0.1709
	layer.dense (128)
	layer.dense (32)
exp_a5	layer.dense (64)	5	dataset2	82.86%	0.1123
	layer.dense (128)
	layer.dense (32)
exp_a6	layer.dense (64)	10	dataset2	82.85%	0.1123
	layer.dense (128)

${}^{\mathrm{a}}$ Experiments are listed in the table considering two k-folds, dense layer and input dataset.${}^{\mathrm{b}}$ There are three hidden layers used in neural network. Example: The neural network for model experiment (exp) is shown in Figure 4.${}^{\mathrm{c}}$ All input datasets are examined with respect to k-fold and input dataset (see 3.3).

Figure 5. Machine learning model flow diagram.

Table 1. Machine learning parameter setting.

Parameters	Values
Input dataset	Six variables and one feature from the logical scenario
Input parameters	(i) velocity (V1), (ii) velocity (V2), (iii) velocity (V3),
	(iv) distance between V2-V1,(v) position V1, (vi) time to lane change
Input variable-	Monte Carlo (Khatun et al., 2021a),
optimization	parameter-based analysis (Khatun et al., 2022)
Dataset1	Total 1122 test case
Dataset2	Total 350 test case
Outcomes/features	Collision detection
Batch size	64
Epoch	2000
Learning rate	0.001
Optimizer	Adam
Dense Layer	3 hidden layers
k-fold	5 and 10

Figure 6. Accuracy and loss of training dataset and validation dataset.

Figure 7. Accuracy and loss of test dataset.

Khatun, M., Caldeira, G. B., Jung, R., & Glaß, M. (2021a). An optimization and validation method to detect the collision scenarios and identifying the safety specification of highly automated driving vehicle. In 21st international conference on control, automation and systems (ICCAS) (pp. 1570–1575). IEEE. https://doi.org/10.23919/ICCAS52745.2021.9649806

 Google Scholar

Khatun, M., Litagin, H., Jung, R., & Glaß, M. (2022). An approach for deriving reduced collision scenarios for highly automated driving systems. In International conference on computer safety, reliability, and security (pp. 166–177). Springer.

 Google Scholar