huanghua1668/dissertation_HuaHuang

Dataset, code, and results of Hua's dissertation

Dataset

Data extraction process:

1. Use extract_lane_change.py to extract lane change sequences from both dataset, and extract_lane_change_merge_after.py to extract lane change abortions.
2. Use extract_sample.py and extract_sample_merge_after.py to extract the exact snapshot from the lane change sequences for the downstream prediction task.
3. Use data_preprocess.py to assemble the extracted samples; split to training and testing; generate OOD samples.

Dataset	Description	Generating program
us80.npz	Raw lane changes from US80, each row is one sample, with record [merge in front/after(0/1), u0, du0, du1, du2, dx0, dx1, dx2, dy0, dy1, dy2, y].	extract_sample.py, extract_sample_merge_after.py, and preprocess_both_dataset() in data_preprocess.py
us101.npz	Raw lane changes from US101, each row is one sample, with record [merge in front/after(0/1), u0, du0, du1, du2, dx0, dx1, dx2, dy0, dy1, dy2, y]	extract_sample.py, extract_sample_merge_after.py, and preprocess_both_dataset() in data_preprocess.py
samples_relabeled_by_decrease_in_dx.npz	Samples with label based on $\Delta x$, each row is a sample with record [index, dt, u0, du0, du1, du2, dx0, dx1, dx2, dy0, dy1, dy2, y]	extract_samples_relabel_as_change_in_distance() in extract_sample.py
combined_dataset.npz	Combine us80 and us101 datasets, do feature selection, and normalize them to have 0 mean and 1 std, in which f['a']=x_train, f['b'']=y_train, f['c']=x_test, f['d']=y_test, f['e']=xGenerated), each sample is of feature x = [dv0, dv1, dx0, dx1]	prepare_validate_and_generate_ood() in data_process.py
combined_dataset_before_feature_selection.npz	Combine us80 and us101 datasets and normalize them to have 0 mean and 1 std, in which f['a']=x_train, f['b'']=y_train, f['c']=x_test, f['d']=y_test), each sample is of feature x = [v_ego, dv0, dv1, dv2, dx0, dx1, dx2, dy0, dy1, dy2]	prepare_validate_and_feature_selection() in data_process.py
combined_dataset_trainUs80_testUs101.npz	us80 as training dataset, us101 as testing dataset. Normalize samples to have 0 mean and 1 std, in which f['a']=x_train, f['b'']=y_train, f['c']=x_test, f['d']=y_test, f['e']=xGenerated), each sample is of feature x = [dv0, dv1, dx0, dx1]	prepare_validate_and_generate_ood_trainUs80_testUs101() in data_process.py
combined_dataset_trainUs101_testUs80.npz	us101 as training dataset, us80 as testing dataset. Normalize samples to have 0 mean and 1 std, in which f['a']=x_train, f['b'']=y_train, f['c']=x_test, f['d']=y_test, f['e']=xGenerated), each sample is of feature x = [dv0, dv1, dx0, dx1]	prepare_validate_and_generate_ood_trainUs80_testUs101() in data_process.py
lane_changes_trajectories.csv	demonstrative lane change trajectories	extract_lane_changes.py
min_dis_within_us80.npz	percentage of minimum distance to other samples in dataset us80	cal_distance() in data_preprocess.py
min_dis_within_us101.npz	percentage of minimum distance to other samples in dataset us101	cal_distance() in data_preprocess.py

Results

Figure/table	Supporting data	Generating program	Plot function (in plot_utils.py)
Figure 3.2: Lane changes	lane_changes_trajectories.csv	extract_lane_changes.py	plot_trajectory()
Figure 3.3: Label based on the change in ∆x.	samples_relabeled_by_decrease_in_dx.npz	extract_samples_relabel_as_change_in_distance() in extract_sample.py	visualize_sample_labeled_by_dx()
Figure 3.4: Accelerations of lag vehicles.	lane_changes_trajectories.csv	extract_lane_changes.py	plot_accelerations()
Figure 3.5: Lane changes in dataset I-80.	us80.npz	preprocess_both_dataset() in data_preprocess.py	plot_ngsim_scatter()
Figure 3.6: Lane changes in dataset US-101.	us101.npz	preprocess_both_dataset() in data_preprocess.py	plot_ngsim_scatter()
Figure 4.1: Confidence map for α = 0 (left) and α = 1 (right).	moons_confidence_alpha0.npz, moons_confidence_alpha1.npz	moons_csnn_alpha_effect_demo.py	plot_confidence_map_alpha_impact()
Figure 4.2: Evolution of the support circles during training.	/evolvement	moons_csnn_evolvement.py	plot_evolvement()
Figure 4.3: Evolution of the confidence during training.	/evolvement	moons_csnn_evolvement.py	plot_evolvement()
Figure 4.4: The radius penalty effect.	/redius_penalty_effect	moons_csnn_radius_penalty_effect.py	plot_radius_penalty_effect()
Figure 4.5: Samples in the moons dataset.	moons_train_test_ood.npz	moons_csnn.py	plot_moons_scatter()
Figure 4.6: Convergence of test accuracy, non-zero output ratio, and AUROC.	moons_acc_nzs_auroc.npz	moons_csnn.py	plot_save_acc_nzs_auroc()
Figure 4.7: Histograms of confidences for in-distribution and OOD samples.	moons_hist_confidence.npz	moons_csnn.py	plot_distribution()
Figure 4.8: ROC curve of the moons dataset	moons_roc.npz	moons_csnn.py	plot_save_roc()
Figure 4.9: Best test set accuracy vs. number of selected features		ngsim_feature_selection.py	plot_feautre_selection_results()
Figure 4.10: Minimum distance to other in-distribution samples.	min_dis_within_us80.npz, min_dis_within_us101.npz	cal_distance() in data_preprocess.py	plot_min_distance_within_dataset()
Figure 4.11: Histograms of confidences on the NGSIM dataset and the generated OOD samples.	ngsim_hist_confidence_epoch240.npz	ngsim_csnn.py	plot_distribution()
Figure 4.12: OOD detection ROC curve for one run of training the CSNN on the NGSIM dataset.	ngsim_roc_epoch240.npz	ngsim_csnn.py	plot_save_roc()
Figure 4.13: Test accuracy and AUROC vs α for the CSNN.	/ngsim_lambda_effect	ngsim_csnn.py	plot_auc_acc_csnn()
Table 4.2: Test accuracy and AUROC in OOD detection, averaged of 10 independent runs. (MLP)	mlp_mean_std_accs_aucs_net4.npz	ngsim_mlp.py
Table 4.2: Test accuracy and AUROC in OOD detection, averaged of 10 independent runs. (DUQ)	/duq	ngsim_duq.py
Table 4.2: Test accuracy and AUROC in OOD detection, averaged of 10 independent runs. (Deep ensemble)	/deep_ensemble	ngsim_deep_ensemble.py
Figure 4.14: Entropy distribution of in-distribution and OOD samples.	/deep_ensemble/ngsim_hist_confidence.npz	ngsim_deep_ensemble.py	plot_distribution()
Figure 4.15: ROC curve obtained with deep ensemble.	/deep_ensemble/ngsim_roc.npz	ngsim_deep_ensemble.py	plot_save_roc()
Figure 4.16: Entropy of average prediction.	moons_confidence_map_deep_ensemble.npz	moons_deep_ensemble.py	plot_moons_de()
Table 4.3: Generalization of network with compact support.	/train_us80_test_us101 and /train_us101_test_us80	ngsim_mlp.py, ngsim_csnn.py, ngsim_duq.py, ngsim_deep_ensemble.py