Detecting Transportation Mode Using Dense Smartphone GPS Trajectories and Transformer Models

Yuandong Zhang; Othmane Echchabi; Tianshu Feng; Wenyi Zhang; Hsuai-Kai Liao; Charles Chang

cs.LGFeb 27, 2026

Detecting Transportation Mode Using Dense Smartphone GPS Trajectories and Transformer Models

Authors: Yuandong Zhang, Othmane Echchabi, Tianshu Feng, Wenyi Zhang, Hsuai-Kai Liao, Charles Chang

Organizations: University of California San Diego, CA, USA · McGill University Mila - Quebec AI Institute Montréal, QC, Canada · University of Pennsylvania Philadelphia, PA, USA · Duke Kunshan University Kunshan, Jiangsu, China

Abstract

Transportation mode detection is an important topic within GeoAI and transportation research. In this study, we introduce SpeedTransformer, a novel Transformer-based model that relies solely on speed inputs to infer transportation modes from dense smartphone GPS trajectories. In benchmark experiments, SpeedTransformer outperformed traditional deep learning models, such as the Long Short-Term Memory (LSTM) network. Moreover, the model demonstrated strong flexibility in transfer learning, achieving high accuracy across geographical regions after fine-tuning with small datasets. Finally, we deployed the model in a real-world experiment, where it consistently outperformed baseline models under complex built environments and high data uncertainty. These findings suggest that Transformer architectures, when combined with dense GPS trajectories, hold substantial potential for advancing transportation mode detection and broader mobility-related research.

Figures & tables

Mode	Data Points		Unique Trips
Mode	MOBIS	Geolife	MOBIS	Geolife
Bike	4,251,028	746,098	40,171	1,555
Bus	4,606,409	1,061,196	74,324	1,847
Car	88,965,473	627,047	542,078	1,293
Train	8,379,962	788,250	130,964	772
Walk	38,463,266	1,215,054	743,425	3,960

Table 1: Comparison of MOBIS and Geolife Datasets

Model	Test Acc. (%)
SpeedTransformer (Ours)	95.97
Deep-ViT [ Ribeiro et al., 2024 ]	92.96
LSTM-based DNN [ Yu, 2020 ]	92.70
LSTM-Attention (Baseline)	92.40
CE-RCRF [ Zeng et al., 2023 ]	85.23
SECA [ Dabiri et al., 2020 ]	84.80

Table 2: Test Accuracy Comparison on Geolife (Ordered by Performance)

Figure 3: Validation accuracies over epochs for Geolife and MOBIS. The SpeedTransformer consistently converges faster and achieves higher overall accuracy than the LSTM-Attention baseline on both datasets.

Figure 4: Per class F1-Score for Geolife and MOBIS trainings using SpeedTransformer and LSTM. The SpeedTransformer consistently achieves better results than the LSTM-Attention across all classes on both datasets.

Mode	Deep-ViT		SpeedTransformer
Mode	Precision (%)	Recall (%)	Precision (%)	Recall (%)
Walk	95.04	96.12	96.72	99.90
Bike	93.79	91.40	98.54	95.45
Bus	89.68	91.13	98.14	95.49
Car	88.96	87.26	87.33	93.65
Train	90.00	79.41	95.66	92.93

Table 3: Validation Precision and Recall Comparison between Deep-ViT and SpeedTransformer on Geolife

Model	100 trips	200 trips
LSTM-Attention	75.47%	79.15%
SpeedTransformer	80.53%	86.13%

Table 4: Model Accuracy during Cross-Dataset Fine-Tuning from MOBIS to Small Geolife Subsets

Panel of Figure 5 — Figure 5: Trip tracking interface in the CarbonClever application: (a) Trip initiation screen, where users start a new trip recording. (b) Active tracking screen showing real-time trip duration and status. (c) Trip completion and mode confirmation screen, where users select and verify the transportation mode. All interface texts are originally in Chinese.

Figure 6: Example of real-world GPS trajectories. The red line indicates the main route, with varying point density reflecting heterogeneous sampling frequencies. Purple circles mark signal interference and positioning noise.

Mode	Number of Data Points	Number of Unique Trips
Walk	28,985	100
Bus	32,547	259
Car	40,885	205
Bike	4,901	36
Train	1,505	49
Total	108,823	649

Table 5: Summary of the Mini-Program Dataset

Data Subset (%)	LSTM-Attention Acc. (%)	SpeedTransformer Acc. (%)
40% (251 trips)	$88.79\pm 2.15$	$90.27\pm 2.04$
50% (314 trips)	$88.82\pm 1.32$	$89.98\pm 2.55$
60% (377 trips)	$89.34\pm 1.86$	$90.40\pm 2.17$
70% (440 trips)	$90.04\pm 2.26$	$91.72\pm 3.27$
80% (502 trips)	$89.30\pm 2.52$	$90.11\pm 4.54$

Table 6: Fine-Tuning Accuracies of Models with Real-World Data Subsets (5-fold CV; mean

\pm

std, % accuracy)

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Mode	Min Speed (km/h)	Max Speed (km/h)
Bike	0.5	50
Bus	1.0	120
Car	3.0	180
Train	3.0	350
Walk	0.1	15

Appendix

Table A1: Speed Thresholds for Each Mode

Mode	Number of Data Points	Number of Unique Trips
Aerialway	109	7
Airplane	718,156	3,185
Bicycle	6,568,587	41,166
Bus	7,450,378	77,361
Car	160,348,026	553,259
Ferry	228	12

Appendix

Table A2: Summary of the MOBIS Dataset

Mode	Number of Data Points	Number of Unique Trips
Airplane	9,196	14
Bike	951,350	1,555
Boat	3,565	7
Bus	1,271,062	1,851
Car	512,939	782
Motorcycle	336	2

Appendix

Table A3: Summary of the Labeled Geolife Dataset

Figure F1: Effect of window size on test accuracy for the Geolife dataset. A window size of 200 provides the optimal balance between temporal context and computational efficiency, achieving the highest accuracy (95.97%).

W. Size	Test Acc. (%)	Test Loss	Run Name
20	77.58	0.5778	Geolife_ws20_lr2e-4_bs512_h8_d128_kv4_do0.1
50	84.86	0.4043	Geolife_ws50_lr2e-4_bs512_h8_d128_kv4_do0.1
100	92.07	0.2600	Geolife_ws100_lr2e-4_bs512_h8_d128_kv4_do0.1
200	95.97	0.1525	Geolife_ws200_lr2e-4_bs512_h8_d128_kv4_do0.1
300	95.04	0.1753	Geolife_ws300_lr2e-4_bs512_h8_d128_kv4_do0.1
400	94.53	0.1703	Geolife_ws400_lr2e-4_bs256_h8_d128_kv4_do0.1

Appendix

Table F12: Test accuracy and loss across different window sizes (Geolife dataset).

Figure G2: Transformer architecture for transportation mode classification.

Panel of Figure 19 — Figure G3: Two GPS speed sequences share the same multiset of speeds but in different orders. Without position information the embeddings for repeated values (e.g., $7.8\,m/s$ ) are identical. With positional encoding we form $\mathbf{z}_{t}=E(v_{t})+P(t)$ , allowing attention to model order-dependent patterns.

Model	Training Setup	Bike	Bus	Car	Train	Walk
LSTM	Geolife	0.92	0.86	0.86	0.92	0.99
	MOBIS	0.61	0.84	0.99	0.28	0.99
	MOBIS $\rightarrow$ Geolife	0.59	0.84	0.57	0.79	0.72
	MOBIS $\rightarrow$ Mini-Program	0.39	0.69	0.93	0.79	0.86
Transformer	Geolife	0.93	0.96	0.92	0.87	0.99
	MOBIS	0.78	0.88	0.98	0.43	0.99

Appendix

Table G13: Per-class accuracy (recall) of LSTM and Transformer models across datasets. Each model was trained on Geolife and MOBIS, and fine-tuned from MOBIS to Geolife and Mini-Program datasets.

Rule	Threshold (m/s)	Description / Condition for Mode Assignment
walk_p95_max	$\leq 3.0$	Windows with 95th-percentile speed below 3.0 m/s are classified as walk .
bike_p95_max	$\leq 8.0$	Windows with 95th-percentile speed between 3.0–8.0 m/s are classified as bike .
road_p95_min	$\geq 17.4$ (18.0*)	Windows exceeding this threshold are treated as motorized road transport ( bus or car ).
rail_p95_min	$\geq 27.5$ (40.0*)	Windows with very high 95th-percentile speed are classified as rail .
stop_thresh	$<0.5$	Speeds below 0.5 m/s are considered “stopped” for computing stop ratio.
bus_stop_ratio_min	$\geq 0.05$	Within the motorized range, higher stop ratios indicate bus .

Appendix

Table H14: Calibrated heuristic rules for transportation mode classification. Thresholds are tuned separately for each dataset using the training split. *A different threshold was used for Geolife

Dataset	Mode	Precision	Recall	F1-score	Support
Geolife	Bike	0.2818	0.3584	0.3155	3,200
	Bus	1.0000	0.0000	0.0000	4,863
	Car	0.0530	0.0871	0.0659	2,251
	Train	0.4256	0.9512	0.5881	4,119
	Walk	0.9463	0.4856	0.6418	5,225
	Macro Avg.	0.5413	0.3765	0.3223

Appendix

Table H15: Performance of the calibrated rule-based model on Geolife and MOBIS datasets.

Figure I4: GPU and memory utilization of the Transformer model during training.

Figure L5: Sampling Frequency Distribution

Note: GQA: Grouped-Query Attention; MHA: Multi-Head Attention; RoPE: Rotary Positional Embeddings; Image Gen.: DeepInsight Transform; Params: approximate number of trainable parameters using the best configuration.
Model	Pre-proc.	Input	Attention	PosEnc.	Params
SpeedTransformer	Speed Calc.	1D Speed	GQA	RoPE	0.73M
Deep-ViT [ Ribeiro et al., 2024 ]	Image Gen.	2D Image	Vision MHA	Learnable	0.53M
TMD-BERT [ Drosouli et al., 2023 ]	Discretisation	Tokens	Bi-Directional MHA	Learnable	110M

Appendix

Table M16: Architectural Comparison of Transformer-based Mobility Models

Panel of Figure 26 — Figure N6: Confusion matrices for SpeedTransformer on the Geolife and MOBIS test sets. Values are raw counts; MOBIS exhibits stronger class imbalance, leading to larger absolute counts.

Figure N7: Per class F1-Score for Fine-tuning on Geolife and our real-world field experiment data using SpeedTransformer and LSTM. The SpeedTransformer consistently achieves better results than the LSTM-Attention across all classes in both tasks.

Panel of Figure 28 — Figure O8: Sequence-level speed dynamics for two dominant train failure profiles. The low_speed_confusion profile shows relatively smooth, gradual ramps; the high_speed_confusion profile has higher variance with abrupt fluctuations and spikes. Shaded regions indicate the 25–75% interquartile range across sequences.

Panel of Figure 29 — Figure O9: Distributional heterogeneity in train-class speed statistics. Mean speed and joint mean–std representations overlap substantially between correct and incorrect train windows, showing that simple aggregates cannot isolate the failure profiles ( low_speed_confusion vs high_speed_confusion ).

Intelligent Transportation Systems

Intelligent Transportation Systems apply sensing, communication, AI, and optimization to make transport networks safer and more efficient.

See all 299 papers

Hyperspectral Trajectory Image for Multi-Month Trajectory Anomaly Detection
Mar 26, 2026Md Awsafur Rahman, Chandrakanth Gudavalli, Hardik Prajapati +1Anomaly Localization Transformer Attention
UC Santa Barbara, USA
TrajTok: Adaptive Spatial Tokenization for Trajectory Representation Learning
May 19, 2026Zhen Xiong, Shang-Ling Hsu, Cyrus ShahabiGeospatial Representation Learning Trajectory Representation Learning
University of Southern California
Non-Parametric Spatiotemporal Trajectory Prediction via State-Conditioned Transition Sampling
Aug 14, 2026Michael Fore, Akshay Jain, Justin Downes +2Trajectory Generation Vessel Trajectory Prediction
Amazon Web Services

Mar 26, 2026J/K move · Enter open · S save

Hyperspectral Trajectory Image for Multi-Month Trajectory Anomaly Detection

Md Awsafur Rahman, Chandrakanth Gudavalli, Hardik Prajapati, B. S. Manjunath

UC Santa Barbara, USA

Anomaly Localization Transformer Attention Time Series Representation Learning Time-Series Anomaly Detection

Open paper AskPDF arXiv

Trajectory anomaly detection underpins applications from fraud detection to urban mobility analysis. Dense GPS preserves fine-grained evidence such as abnormal speeds and short-duration events, but its quadratic cost makes multi-month analysis intractable; sparse stay-point methods scale by discarding that evidence and require a separate modeling regime. We argue that this bottleneck is unnecessary: dense and sparse trajectories share a natural two-dimensional cyclic structure along within-day and across-day axes. TITAnD (Trajectory Image Transformer for Anomaly Detection) is the first framework to cast both in a single representation, a Hyperspectral Trajectory Image (HTI), a day

\times

time-of-day grid whose channels encode spatial, semantic, temporal, and kinematic information. Under this formulation, agent-level detection reduces to image classification and temporal localization to semantic segmentation. The Cyclic Factorized Transformer (CFT) models the two temporal axes directly, reducing attention cost by up to two orders of magnitude. With the HTI, multi-month dense anomaly detection becomes feasible for the first time, and CFT keeps it accurate and fast. Empirically, TITAnD matches or beats every evaluated baseline in AUC-PR on all four benchmarks. It surpasses vision models such as U-Net while using as few as 6.5M parameters, and it runs 11--75

\times

faster than a capacity-matched flat Transformer on the same trajectory images.

Abstract

Intelligent Transportation Systems

Explore similar work

Hyperspectral Trajectory Image for Multi-Month Trajectory Anomaly Detection

TrajTok: Adaptive Spatial Tokenization for Trajectory Representation Learning

Non-Parametric Spatiotemporal Trajectory Prediction via State-Conditioned Transition Sampling