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

cs.LGAug 18, 2026

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

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

Organizations: Department of Computer Science and Engineering, Indian Institute of Technology Indore, Khandwa Road, Simrol, Indore, 453552, Madhya Pradesh, India · Department of Civil Engineering, Indian Institute of Technology Indore, Khandwa Road, Simrol, Indore, 453552, Madhya Pradesh, India · Civil Engineering Program, Ingram School of Engineering, Texas State University, RFM 5202, San Marcos, 78666, TX, USA

Abstract

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. We address this gap by introducing a comprehensive dataset consisting of over 129,000 labeled multivariate time-series feature windows, gathered across 153 simulator rides from 51 participants under No, Low, and High TP scenarios. Across each sequence, we capture 64 distinct attributes covering vehicle motion, rider control actions, spatial proximity, and rule compliance indicators. Using this dataset, we introduce MotoSafety, a new edge-AI framework built on the Learned Temporal Importance (LTI) concept. 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.

Figures & tables

Component	Details
Platform	Fixed-base PTW frame with ISO certification
Controls	Acceleration, clutch, hydraulic brake unit, 5-speed shifting mechanism, handlebar steering, parking brake
Sensors	Wire-wound servo potentiometers (type 50WW):
	Resistance: 5 k $\Omega$ (±10%); Linearity error: ±0.5% (IEC 393 compliant), Angular range: 355° ± 3°; Full mechanical rotation: 360°
	Maximum power: 3 W (at 70°C); Expected operational life: 2 million cycles
	Safe operating limits: –40°C to +105°C

Table 1: Two-wheeler simulator: key technical parameters

Panel of Figure 3 — Figure 3: Participants performing the simulator test.

Panel of Figure 4 — Figure 4: Sample snapshots of the simulated riding environment.

Attribute	Category / Range	Mean (SD) or %
Age (years)	18–42	26.4 (5.7)
Riding experience (years)	$\geq 2$	5.6 (3.1)
License status	Valid PTW license	100%
Education level	Graduate / B.Tech	58.8% / 41.2%
Simulator exposure	No / Yes	92.2% / 7.8%
Health status	Medically fit	100%

Table 2: Demographic and riding profile of participants (N = 51)

Figure 5: Flow of dataset collection and experimental design.

Feature Category	Included Parameters	Count
Vehicle Controls	Ignition status, Engine state, Throttle position, Brake application, Clutch engagement, Handbrake, Steering angle, Gear position, Headlight usage, Horn activation	10
Vehicle Performance	Speed, Engine RPM, Fuel consumption, Distance covered	4
Lighting & Indicators	Indicator signal, Pre-move indication, Junction turning signal, Lane-change signal, Headlight non-compliance	5
Behavioral Violations	Speeding, Incorrect junction/intersection speed, Speed violations on speed breakers, Gap maintenance errors, Hazardous overtaking, Unsigned turns, Lane misuse, Driving on the wrong side, Handbrake use while moving, Riding the clutch, Faulty gear shift order, Improper clutch release, Shifting without clutch engagement, Proper gear selection before start, Gradual clutch engagement	15
Traffic Infractions	White line violations, Yellow line crossings, Stop line breaches, Signal jumping, Entering no-entry zones, Illegal U-turns, Parking infringements	7
Temporal Context	Time stamp, Time Pressure condition (0=HTP, 1=LTP, 2=NTP)	2

Table 3: Overview of simulator-derived features and their categories

Figure 6: The proposed MotoSafety architecture.

Module	Complexity
Dilated CNN encoder (×3 branches)	$\mathcal{O}(L\cdot k\cdot C)$
Bi-LSTM (2 layers)	$\mathcal{O}(L\cdot H^{2})$
TIP scoring + weighted sum	$\mathcal{O}(L\cdot H)$
Gated fusion + SE block	$\mathcal{O}(1)$ (fixed-size input)
Multi-head attention (post-pooling)	$\mathcal{O}(1)$ (fixed-size input)
MotoSafety total	$\mathcal{O}(L)$

Table 4: Per-module computational complexity.

L

: sequence length,

C

: feature channels,

H

: hidden dimension,

k

: kernel size.

Classification Model
Component	Setting
Input sequence length	96 timesteps (960 ms)
Feature dimension	64
CNN hidden channels	32 per branch, 64 fused
BiLSTM hidden dimensions	128 (bidirectional)
Total model parameters	$\sim$ 1.15 million

Table 5: Training configuration for MotoSafety classification and forecasting models.

Feature	NTP	LTP	HTP
Speed (km/h)	33.12 $\pm$ 19.42	39.69 $\pm$ 21.67	49.00 $\pm$ 26.48
Sudden braking	1.32 $\pm$ 2.54	1.38 $\pm$ 2.36	1.80 $\pm$ 2.01
Dangerous overtaking	1.07 $\pm$ 1.18	1.21 $\pm$ 1.27	1.36 $\pm$ 1.52
Headway distance (m)	8.99 $\pm$ 14.38	8.20 $\pm$ 14.34	7.75 $\pm$ 13.88

Table 6: Behavioral differences across TP: NTP/LTP/HTP (Mean

\pm

Std).

Models	MotoSafety		Time-LLM		iTransformer		PatchTST
Metric	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
96	0.033 $\pm$ 0.005	0.082 $\pm$ 0.007	0.136	0.148	0.161	0.237	0.188	0.208
192	0.037 $\pm$ 0.006	0.086 $\pm$ 0.008	0.145	0.153	0.170	0.246	0.205	0.223
336	0.041 $\pm$ 0.006	0.097 $\pm$ 0.010	0.192	0.215	0.177	0.258	0.227	0.272
720	0.045 $\pm$ 0.007	0.111 $\pm$ 0.012	0.212	0.293	0.183	0.262	0.267	0.338
Avg	0.039	0.094	0.171	0.210	0.173	0.251	0.222	0.260

Table 7: Multivariate long-term forecasting results (mean

\pm

std over 5 runs). Input lookback length

L=96

. Forecast horizons

H\in\{96,192,336,720\}

. Avg is averaged over all four prediction horizons.

Model	Acc. (%)	Prec.	Rec.	F1-Score (%)	AUC (%)
	Mean $\pm$ Std	Result	Result	Mean $\pm$ Std	Mean $\pm$ Std
RF	$84.93\pm 0.16^{\ast\ast\ast}$	85.24	84.87	$85.02\pm 0.23^{\ast\ast\ast}$	$93.68\pm 0.31^{\ast\ast\ast}$
CNN	$90.19\pm 0.11^{\ast\ast\ast}$	93.82	80.68	$86.74\pm 0.21^{\ast\ast\ast}$	$97.62\pm 0.22^{\ast\ast\ast}$
RNN	$90.94\pm 0.12^{\ast\ast\ast}$	83.41	96.38	$89.43\pm 0.32^{\ast\ast\ast}$	$97.27\pm 0.34^{\ast\ast\ast}$
LLM4TS	$90.06\pm 0.14^{\ast\ast\ast}$	90.58	83.23	$86.81\pm 0.31^{\ast\ast\ast}$	$97.31\pm 0.42^{\ast\ast\ast}$
Informer	$93.61\pm 0.07^{\ast\ast\ast}$	89.72	94.88	$92.21\pm 0.22^{\ast\ast}$	$99.12\pm 0.21^{\ast\ast}$

Table 8: Comparison of model performance (mean

\pm

standard deviation over 5 runs). We used the paired Wilcoxon signed-rank test with Holm–Bonferroni adjustment to determine statistical significance (

\alpha=0.05

). Significance is denoted by:

{}^{***}p<0.001

{}^{**}p<0.01

{}^{*}p<0.05

. All baseline models were compared against the proposed MotoSafety framework.

Horizon $H$	Time (s)	MSE	MAE
96	0.96	0.033 $\pm$ 0.005	0.082 $\pm$ 0.007
192	1.92	0.037 $\pm$ 0.006	0.086 $\pm$ 0.008
336	3.36	0.041 $\pm$ 0.006	0.097 $\pm$ 0.010
720	7.20	0.045 $\pm$ 0.007	0.111 $\pm$ 0.012

Table 9: Forecasting performance by prediction horizon (mean over 5 runs)

Model	Accuracy (%)	AUC (%)
Without TP Feature	$94.09\pm 0.04$	$99.10\pm 0.01$
With Predicted TP Feature	$94.82\pm 0.02^{***}$	$99.24\pm 0.01^{***}$
With Ground Truth TP Feature	$\mathbf{94.97\pm 0.01^{***}}$	$\mathbf{99.33\pm 0.01^{***}}$

Table 10: Performance comparison of MotoSafety showing improvement from adding TP features. Significance markers indicate improvement over the Without TP baseline:

{}^{***}p<0.001

True / Predicted	0: HTP	1: LTP	2: NTP
0: HTP	8878	229	163
1: LTP	76	7326	568
2: NTP	82	1661	6859

Table 11: Confusion matrix (raw counts) for TST-based TP prediction.

Figure 7: Confusion matrix of MotoSafety predictions on the main PTW simulator dataset under TP.

Dataset	Details	Safety Domain	Application	Country	MST	Prior Work Accuracy
Proposed Dataset	PTW Simulator with GT TP	PTW Safety	Collision Prediction	India	94.97	–
[ 41 ]	PTW High-Fidelity Simulator	PTW Safety	Collision Prediction	Germany	99.49	91.0 [RF, GB [ 42 ] ]
[ 4 ]	PTW Real Time Fall Event Data	PTW Safety	Fall Detection	France	96.10	91.59 [DT [ 8 ] ]
[ 40 ]	Human Activity Recognition	Human Safety	Activity Recognition	Italy	97.66	83.35 [K-means + NB [ 14 ] ]
[ 50 ]	Wearable-Sensor Movement	Human Clinical Safety	Exercise Quality Assessment	Turkey	99.65	88.65[MTMM-DTW [ 49 ] ]

Table 12: Cross-Safety-Domain Evaluation of MotoSafety Accuracy (%) and Prior Work Accuracy (%). MST: MotoSafety.

Variant	Accuracy (%)	$\Delta$ Acc.
Full Model	94.97	–
No BiLSTM	90.42	$-4.55$
No SE Block	91.24	$-3.73$
No Attention	90.78	$-4.19$
No CNN Branch	92.66	$-2.31$
No Gated Fusion	92.90	$-2.07$

Table 13: Ablation Study of MotoSafety Components.

Pooling Strategy	Accuracy (%)	AUC (%)
Mean-pooling	92.23 $\pm$ 0.15	98.80 $\pm$ 0.22
Max-pooling	93.41 $\pm$ 0.12	99.11 $\pm$ 0.18
Last-timestep	92.87 $\pm$ 0.14	98.97 $\pm$ 0.20
TIP	94.97 $\pm$ 0.08	99.33 $\pm$ 0.30

Table 14: Comparison of MotoSafety with TIP against standard fixed temporal-pooling strategies (mean

\pm

std over 5 runs).

Features	Accuracy (%)
3	80.22
5	83.11
7	88.78
9	90.16
21	93.91
All 64 (simulator)	94.97

Table 15: MotoSafety accuracy with real-bike features (IMU+GPS).

A. MotoSafety Efficiency & Deployability
Metric	Value	Significance
Total Parameters	1,149,048	Low Complexity
Model Size	4.38 MB	Edge-ready Storage
Throughput	$>7,400$ s/sec	High-speed Inference
Arch. Complexity	$\mathcal{O}(L)$	Linear Scalability
Edge-AI Deployability	Yes	Suitable for wearable

Table 16: MotoSafety Efficiency Metrics and Comparative Latency.

Intelligent Transportation Systems

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

See all 299 papers

MOTOR: A Multimodal Dataset for Two-Wheeler Rider Behavior Understanding
May 21, 2026Varun A. Paturkar, Shankar Gangisetty, C. V. JawaharHuman Activity Recognition Video Action Recognition
CVIT, IIIT-Hyderabad, India
Progressive Risk Estimation for Accident Anticipation
Sep 26, 2026Samet Hicsonmez, Eray Çakar, Nermin Samet +1Traffic Accident Anticipation Autonomous Driving
Independent Researcher, Paris, France · Department of Computer Engineering, KUIS AI, Koç University, Istanbul, Turkey · Valeo.ai, Paris, France
Which Trajectory Is Safer? Combining Rule and Learning for Safety Evaluation
Oct 13, 2026Xuewen Luo, Linlian Jiang, Bo Yu +3Autonomous Driving Safety Evaluation
Xuewen Luo and Bo Yu are with the University of Utah. · Linlian Jiang, Xinxin Zuo, and Yang Wang are with Concordia University. · Fan Ding is with Monash University.

May 21, 2026J/K move · Enter open · S save

MOTOR: A Multimodal Dataset for Two-Wheeler Rider Behavior Understanding

Varun A. Paturkar, Shankar Gangisetty, C. V. Jawahar

CVIT, IIIT-Hyderabad, India

Human Activity Recognition Video Action Recognition

Open paper AskPDF arXiv

Two-wheelers account for a disproportionately high share of road fatalities in the Global South. Research on two-wheeler rider behavior, however, lags far behind four-wheelers, where multimodal datasets have driven major advances in Advanced Driver Assistance Systems (ADAS). To address this gap, we present the MOtorized TwO-wheeler Rider (MOTOR) dataset, the first large-scale, multi-view, multimodal resource dedicated to two-wheelers in dense, unstructured traffic. MOTOR comprises 1,629 sequences (25+ hours of video data) collected from 16 riders and integrates synchronized front, rear, and helmet videos, rider eye-gaze from wearable trackers, on-road audio, and telemetry (GPS, accelerometer, gyroscope). Rich annotations capture traffic context, rider state, 12 riding maneuvers spanning conventional and unconventional behaviors, and legality labels (Legal, Illegal, Unspecified). We benchmark rider behavior recognition and maneuver legality classification using state-of-the-art video action recognition backbones (CNN and Transformer-based), extended with multimodal fusion, and find that combining RGB, gaze, and telemetry consistently yields the best performance. MOTOR thus provides a unique foundation for advancing safety-critical understanding of two-wheeler riding. It offers the research community a benchmark to develop and evaluate models for behavior analysis, legality-aware prediction, and intelligent transportation systems. Dataset and code is available at https: //varuniiith.github.io/MOTOR-Dataset/

Abstract

Intelligent Transportation Systems

Explore similar work

MOTOR: A Multimodal Dataset for Two-Wheeler Rider Behavior Understanding

Progressive Risk Estimation for Accident Anticipation

Which Trajectory Is Safer? Combining Rule and Learning for Safety Evaluation