PaperlayerView exact revision on arXiv

Source-linked AI summary

MotoSafety: Edge-AI with Learned Temporal Importance for Two-Wheeler Collision Risk Assessment Under Time Pressure

Sumit S. Shevtekar, Chandresh K. Maurya, Gourab Sil, Subasish Das

arXiv:2608.17823v1cs.LGcs.AIcs.HC

TL;DR

Time Pressure is rarely integrated into research on unsafe two-wheeler riding. This paper introduces MotoSafety, an edge-AI architecture for collision-risk assessment, achieving 94.97% accuracy and 0.039 MSE for forecasting.

  • Problem

    Time Pressure is rarely integrated into research on unsafe driving behavior, leaving limited evidence on two-wheeler collision risk under pressure.

  • Method

    The study builds on a labeled riding dataset and proposes MotoSafety, a deep-learning architecture grounded in the Learned Temporal Importance principle.

  • Results

    MotoSafety outperformed baselines for collision-risk classification and long-term forecasting, achieving 94.97% accuracy and 0.039 MSE, respectively.

  • Takeaways & Limitations

    The proposed framework advances two-wheeler collision-risk assessment within a lightweight edge-AI approach.

  • Takeaways & Limitations

    The study indicates that integrating contextual information remains future work to enhance robustness.

Abstract

from arXiv · show

Powered two-wheeler riders face critical safety challenges in low- and middle-income countries, yet limited studies exist on how cognitive stressors such as Time Pressure influence collision risk. To address this gap, we introduce a large-scale dataset of over 129,000 labeled multivariate time-series sequences from 153 simulator rides by 51 participants under No, Low, and High TP, capturing 64 features across vehicle dynamics, control inputs, proximity, and behavioral violations. Building on this dataset, we propose MotoSafety, a novel edge-AI architecture grounded in the Learned Temporal Importance principle. MotoSafety achieves 94.97% accuracy and 99.33% ROC AUC, outperforming ten baselines, including TimesNet and LLM4TS, and achieves 0.039 MSE and 0.094 MAE for forecasting (4.4x lower error than Time-LLM and iTransformer). With only 1.15M parameters and 0.135 ms latency, it is suitable for edge deployment on low-cost CPU hardware. Using ground truth TP as an inductive bias improves accuracy from 94.09% to 94.97%, while predicted TP achieves 94.82%. Using only 21 IMU+GPS features, it achieves 93.91% accuracy, indicating practical deployment. Beyond PTW safety, the architecture shows better transferability to human activity (97.66%) and clinical (99.65%) domains. This lightweight framework advances PTW collision risk assessment, supporting the Safe System Approach for Intelligent Transportation Systems.

1. Introduction

The paper addresses the underexplored role of Time Pressure in powered two-wheeler collision risk, proposing MotoSafety as a lightweight, edge-deployable multivariate time-series model. It contributes a large-scale TP simulator dataset and reports strong classification, forecasting, and deployment performance.

  • Motivation: PTW riders are disproportionately vulnerable, accounting for approximately 21% of global road traffic fatalities and nearly 38% in LMICs.Limited protection, complex traffic, and socioeconomic pressures contribute to their risk.
  • Problem: Human error remains a dominant contributor to PTW crashes, while TP-induced cognitive stress is insufficiently understood in PTW-driving situations.Risky behaviors include overspeeding, abrupt maneuvering, inconsistent braking, and delayed hazard response.
  • Research Gap: Existing crash-prediction approaches inadequately model PTW dynamics, unstable control characteristics, fine-grained temporal behavior, and cognitive stress factors such as TP.Prior work focuses primarily on four-wheelers or general traffic datasets.
  • Contributions: The study collects 129,000 labeled multivariate time series from 51 participants under No, Low, and High TP, using 64 features spanning vehicle dynamics, control inputs, distances, and behavioral violations.The dataset is designed to study TP effects on PTW riders.

4. TP Inductive Bias, Real-World Feasibility, and Architecture Transferabil-

MotoSafety benefits from explicit time-pressure prediction, remains accurate with only 21 IMU+GPS features, and transfers effectively to human-activity and clinical-exercise domains. Hardware validation is identified as future work.

  • Real-World Feasibility: 93.91% accuracy using only 21 IMU+GPS features indicates practical deployment potential.The reduced feature set uses real-world sensor inputs.
  • Real-World Feasibility: Hardware validation remains future work.Practical deployment potential is reported, but hardware validation has not yet been completed.
  • Architecture Transferability: 97.66% accuracy is achieved in human activity and 99.65% in clinical exercise after retraining from scratch.These results demonstrate transferability beyond PTW safety.

2. Related Work

Prior work links time pressure to altered riding behavior and elevated crash risk, while existing risk models rarely incorporate it as a latent precursor, especially for PTWs. This motivates simulator-based multivariate modeling to detect risk earlier under cognitively stressed conditions.

  • Research Gap: Time pressure acts as an upstream cognitive trigger for unsafe outcomes, but it is rarely integrated into collision-risk assessment as a latent risk factor.This gap is particularly important for PTWs because of their narrow safety margins, unstable dynamics, and dependence on rider motor control.
  • Time Pressure and PTW Risk: 181% crash likelihood under high time pressure and 127% under low time pressure were reported relative to no time pressure.Time pressure also narrows accepted gaps and increases speed, control variability, unstable steering, and abrupt braking.
  • Simulation-Based Research: Driving simulators enable safe, high-resolution analysis of collision risk and kinematic forecasting under cognitively demanding conditions such as time pressure.Simulator studies show relative validity for speed selection, braking intensity, and control variability, although absolute behavioral magnitudes can differ from naturalistic settings.
  • Risk Modeling Literature: Recent sequence models capture long-range driving dynamics, but most existing approaches detect externally observable collisions or near-crashes after risk has manifested.Classical machine-learning methods include Random Forests, Support Vector Machines, and feed-forward neural networks, while newer models include Temporal Transformers, Informer, and TimesNet.
  • Study Motivation: The present work addresses this gap by inferring collision risk from high-resolution behavioral dynamics under cognitive stress, supporting high-risk-state detection and intelligent rider-assistance systems.The framing aligns with the Safe System Approach and Vision Zero goals for safer PTW mobility.

3. Methodology

The methodology combines a high-fidelity, ISO-compliant two-wheeler simulator with a controlled urban route and graduated time-pressure conditions to study rider behavior and collision risk. Data from 51 male riders produced 153 sessions and 129,209 labeled overlapping time-series windows built from 64 normalized features.

  • Simulator Setup: The ISO-compliant simulator combines an instrumented motorcycle frame, operational controls, immersive roadway displays, precision sensors, and a 4-actuator motion system.The setup supports controlled investigation of rider behavior, cognitive load, and collision risk under realistic operational conditions.
  • Scenario Design: A 4.8 km urban scenario with a 50 km/h speed limit includes lane configurations, pedestrian crossings, overtaking, intersections, following segments, and dynamic interaction triggers.These events elicit real-time decision-making, adaptive control, and risk-taking under varying time pressure.
  • Time-Pressure Protocol: Three conditions operationalize graduated cognitive load: NTP allows ample time, LTP restricts duration to 90%, and HTP restricts it to 80% with urgency prompts.The standardized examination-deadline scenario is designed to mimic emergency commuting and high-arousal urban-transit stress.
  • Time-Pressure Rationale: Including LTP captures early behavioral degradation between NTP and HTP, improving sensitivity to rising stress and supporting earlier safety interventions.LTP reveals subtle signs such as mild overspeeding and minor lane deviations before riders reach the high-risk HTP state.

Classification Task (Same-Window Risk Detection): Given a 960 ms observation

The classification task uses a 960 ms observation window to determine whether a collision occurs within that same window. Because some positive windows overlap impact, sliding, or aftermath phases, it measures same-window risk separability rather than fixed-lead-time anticipation.

  • Task definition: A 960 ms observation window is used to predict whether a collision event occurs within that window.
  • Same-window overlap: Some positive windows overlap the collision event itself, including impact, sliding, or aftermath phases.
  • Task characterization: The task is same-window collision risk detection rather than fixed-lead-time forecasting.
  • Evaluation scope: It evaluates whether pre-collision and in-progress behavioral signatures are jointly separable from normal riding, not how far in advance a collision can be anticipated.

Forecasting Task (Prospective Kinematic Prediction): Given a lookback win-

MotoSafety performs prospective multi-horizon forecasting of future kinematic states from a 96-timestep lookback window, with horizons extending to 7.2 seconds. This forecasting task is trained separately from collision classification and targets sensor-state trajectories rather than future collision probability.

  • Forecasting Task: L = 96 timesteps are used to predict future kinematic states up to H ∈{96, 192, 336, 720} timesteps ahead.Targets include speed, lean angle, and longitudinal force.
  • Forecasting Task: The task is prospective forecasting with an explicit lead time, but its target is future sensor state rather than future collision probability.It therefore anticipates kinematic precursors of hazardous states while remaining distinct from direct collision-probability forecasting.
  • Forecasting Module: A separate but architecturally similar forecasting network uses CNN and BiLSTM branches, TIP, gating, SE-Block, and MHA to predict future target values.Classification and forecasting share the L = 96 lookback window but differ in prediction target and are trained separately.
  • Forecasting Module: The forecasting module is trained by minimizing Mean Squared Error over the predicted future sequence.The target at each future step is the true kinematic value y_i,L+k.

4. Experimental Setup and Baselines

The experimental setup validates simulator-generated labels, benchmarks MotoSafety against conventional and modern sequence models, and evaluates it under controlled training, splitting, and statistical-testing procedures.

  • Label Validation: 2,000 randomly selected segments from the 129,000-sequence dataset were independently reviewed by two road-safety and human-factors experts for high-risk episode onset and severity.Annotations covered abrupt swerves, loss of balance, and critical near-misses using kinematic thresholds and behavioral cues.
  • Label Validation: κ = 0.87 indicated almost perfect inter-annotator agreement, validating the simulator-generated labels as robust ground truth for model training.Agreement was measured using Cohen’s Kappa coefficient.
  • Baselines: MotoSafety was benchmarked against RF, CNN, RNN, Informer, iTransformer, TimesNet, PatchTST, Time-LLM, and LLM4TS (GPT-2).The baselines span conventional machine learning, deep learning, Transformer, frequency-analysis, patching-based, and fine-tuned LLM architectures.
  • Training Configuration: Training used PyTorch, AdamW, Focal Loss for classification, MSE loss for forecasting, Mixup, EMA, and MC Dropout with 5 samples.Experiments ran for 50 epochs with batch size 64 on an NVIDIA T400 GPU with 4GB VRAM.
  • Evaluation Protocol: An 80/10/10 participant-wise split kept all sessions from each rider together, preventing data leakage.Performance was measured with Accuracy, F1-Score, and ROC-AUC for classification and MSE and MAE for forecasting.
  • Statistical Analysis: Statistical significance was assessed using paired Wilcoxon signed-rank tests over 5 runs with Holm–Bonferroni correction at α = 0.05.Forecasting results used mean absolute error (MAE) alongside MSE.

5. Results

MotoSafety demonstrates strong PTW collision-risk classification and forecasting performance under time pressure, while preserving calibration, transferability, and edge efficiency. Ablations and reduced-feature tests further support the architecture’s learned temporal weighting and practical deployment.

  • Forecasting: 0.039 MSE and 0.094 MAE yield a 4.4× error reduction over Time-LLM (0.171) and iTransformer (0.173), with meaningful signal through H = 720.Across horizons, MSE increases from 0.033 at H = 96 to 0.045 at H = 720 rather than degrading sharply.
  • Collision-risk classification: 94.97% accuracy, 93.7% F1-score, and 99.33% ROC AUC outperform ten baselines, including TimesNet, PatchTST, iTransformer, Time-LLM, and LLM4TS.The high ROC AUC supports low-false-alarm collision-risk assessment for Advanced Rider Assistance Systems.
  • Time-pressure features: 94.97% accuracy with ground-truth TP, 94.82% with predicted TP, and 94.09% without TP show that TP features improve collision-risk prediction.The predicted-versus-oracle gap is 0.15%, supporting use when ground-truth TP is unavailable.
  • Ablation analysis: 94.97% full-model accuracy falls by 4.55% without BiLSTM, 4.19% without MHA, 2.31% without CNN, 2.07% without Gated Fusion, and 3.73% without SE.These ablations support the contribution of long-range dependency modeling, multi-scale extraction, and channel-wise recalibration.

6. Conclusion

MotoSafety combines a large time-pressure two-wheeler dataset with a lightweight learned-temporal-importance architecture for accurate collision-risk assessment and forecasting. Its edge-deployment promise is tempered by simulator, demographic, physiological, and real-world validation limitations.

  • Contributions: The dataset contains over 129,000 labeled multivariate time-series sequences from 153 simulator rides involving 51 participants across no, low, and high time pressure.Each sequence captures 64 features spanning vehicle dynamics, control inputs, proximity, temporal context, and behavioral violations.
  • Performance: 94.97% accuracy and 99.33% ROC AUC were achieved for collision risk assessment, outperforming ten baselines including TimesNet, PatchTST, iTransformer, Time-LLM, and LLM4TS.MotoSafety is grounded in the learned temporal importance principle.
  • Performance: 0.039 MSE and 0.094 MAE were achieved for long-term forecasting, with 4.4× lower error than Time-LLM and iTransformer.The model uses 1.15 million parameters and has 0.135 ms inference latency.
  • Deployment and transferability: 97.66% on human activity recognition and 99.65% on clinical exercise monitoring demonstrate improved transferability beyond powered two-wheeler safety.The architecture has 1.15M parameters, 4.38 MB size, and 0.135 ms latency, supporting potential cloud-independent edge deployment on handlebar devices or smart helmets.
  • Limitations and future work: The study is limited by an entirely male cohort, controlled simulator data, absent physiological stress measures, and the need for prospective operational and hardware validation.Future work will test diverse populations and road environments, integrate multimodal signals, develop adaptive ITS interventions, and apply transfer learning across vehicle types, regions, and cultures.

Data and Code Availability

The authors state that all data and code used in the study will be made available from the corresponding author upon reasonable request after publication.

  • All study data and code will be available from the corresponding author upon reasonable request after publication.

CRediT Authorship Contribution Statement

The CRediT statement assigns Sumit S. Shevtekar primary contributions across conceptualization, data, analysis, methodology, software, validation, visualization, and drafting. Drs. Chandresh K. Maurya, Gourab Sil, and Subasish Das contributed to methodology, supervision, validation, and manuscript review, with additional research, funding, and project roles specified individually.

  • CRediT Authorship Contribution Statement: Sumit S. Shevtekar contributed to conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, and manuscript drafting and review.His listed writing roles include Writing – original draft and Writing – review & editing.
  • CRediT Authorship Contribution Statement: Dr. Chandresh K. Maurya contributed to conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, supervision, validation, and manuscript review.His writing role is Writing – review & editing.
  • CRediT Authorship Contribution Statement: Dr. Gourab Sil contributed to conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, resources, supervision, validation, and manuscript review.His writing role is Writing – review & editing.
  • CRediT Authorship Contribution Statement: Dr. Subasish Das contributed to methodology, supervision, validation, and manuscript review and editing.His listed writing role is Writing – review & editing.

Funding

The work was supported by the IIT Indore Young Faculty Research Catalyzing Grant (YFRCG) Scheme, Project ID IITI/YFRCG/2023-24/01.

  • Funding: The study received support from the IIT Indore Young Faculty Research Catalyzing Grant (YFRCG) Scheme.Project ID: IITI/YFRCG/2023-24/01.
Loading 2608.17823v1…