🤖 AI

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

This paper introduces MotoSafety, a lightweight edge-AI architecture leveraging learned temporal importance to achieve high-accuracy, low-latency collision risk assessment for powered two-wheelers under time pressure, validated by a large-scale simulator dataset and demonstrating strong performance across safety, activity, and clinical domains.

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

Published 2026-08-19

📖 6 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, millions of people navigate the world's busiest streets on two wheels, balancing on a thin line between mobility and danger. In many countries, these riders face a unique set of challenges: narrow lanes, unpredictable traffic, and the intense pressure to arrive quickly. This urgency, known as time pressure, is more than just a feeling of being late; it is a cognitive stressor that narrows a rider's focus, speeds up their decision-making, and often pushes them toward risky behaviors like speeding or sudden braking. While engineers and safety experts have long understood that stress leads to accidents, the precise moment when a rider's mind shifts from safe to dangerous has remained difficult to pinpoint. Traditional safety systems often react only after a crash has happened or rely on simple rules that cannot account for the complex, split-second changes in human behavior.

To bridge this gap, a team of researchers has developed a new way to listen to the story a motorcycle tells before a collision occurs. By studying how riders behave under different levels of stress, they created a digital system capable of recognizing the subtle signs of danger in real-time. This work does not just predict accidents; it learns to understand the rhythm of a ride, identifying the specific moments when a rider's control begins to slip due to the pressure of the clock. The result is a lightweight, fast-thinking tool designed to fit on the motorcycle itself, offering a chance to warn riders before a mistake becomes a tragedy.

The researchers began by building a large library of data that captures the full experience of riding under pressure. They invited fifty-one experienced riders to a high-fidelity simulator, a safe environment that replicates the feel of a real motorcycle with moving parts and a wide, immersive view of the road. Each rider completed three separate journeys: one where they had plenty of time, one where they were slightly rushed, and one where they were under extreme time pressure, racing against a ticking clock to reach a destination. During these rides, sensors recorded sixty-four different details every second, from the speed of the engine and the force applied to the brakes to the distance between the bike and other vehicles. This process generated over 129,000 distinct snapshots of riding behavior, creating a massive dataset that shows exactly how a rider's actions change as stress levels rise.

From this data, the team built a new kind of artificial intelligence called MotoSafety. Unlike older systems that might treat every second of a ride as equally important, this new system learns to pay attention to the moments that matter most. Imagine a system that can sift through hours of video and instantly spot the split second a driver looks the wrong way; MotoSafety does something similar for time. It scans the stream of sensor data and learns to weigh the importance of each moment, focusing intensely on the brief windows where a rider's control becomes unstable. This ability to prioritize critical moments allows the system to spot the early warning signs of a crash, such as erratic steering or sudden, sharp braking, long before a collision actually happens.

The system was tested against ten other advanced methods, including complex models that try to predict the future by looking at vast amounts of historical data. MotoSafety proved to be significantly more accurate, correctly identifying collision risks in nearly 95 percent of cases. It also excelled at forecasting the future path of the motorcycle, predicting where the bike would be in the next few seconds with far less error than its competitors. Crucially, the system is designed to be small and fast. It requires very little computing power and can make a decision in a fraction of a millisecond, which is fast enough to run on a simple processor mounted directly on the motorcycle's handlebars. This means it does not need to send data to a distant cloud server to work; it can think and react instantly, right where the rider is.

The researchers also discovered that knowing the rider's stress level directly improves the system's accuracy. When the system was given the actual time pressure condition of the ride, its performance improved slightly, but even when it had to guess the stress level based on the rider's behavior alone, it remained highly accurate. This suggests that the system can learn to detect the invisible weight of time pressure just by watching how a rider handles the bike. Furthermore, the system works well even when using only a basic set of sensors, such as those found on standard motorcycles today, which measure speed, acceleration, and position. This finding indicates that the technology is ready for real-world use without needing expensive, specialized equipment.

The study also looked beyond motorcycles to see if this approach could help in other areas. When the same system was trained on data from human movement and clinical exercise, it performed with similar high accuracy, suggesting that the method of learning to spot critical moments is a powerful tool for understanding human behavior in many different contexts. However, the researchers are clear that this work is a significant step forward, not a final solution. The data came from a controlled simulator, and while the results are promising, the system still needs to be tested on real roads with real traffic, weather, and fatigue. The team plans to move forward with testing the system on physical devices and expanding the study to include riders of all genders, ensuring that this new layer of safety can protect everyone who rides.

In a world where time is often the enemy of safety, this research offers a new perspective: by understanding how time pressure changes the way we move, we can build systems that help us stay safe. MotoSafety does not stop the clock, but it gives riders a chance to see the danger before it arrives, turning a split-second decision into a moment of awareness. As cities grow more crowded and the pressure to move faster increases, tools like this may become essential, providing a quiet, intelligent guardian for the vulnerable millions who share the road on two wheels.

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

1. Problem Statement

Powered two-wheeler (PTW) riders face disproportionately high fatality rates, particularly in Low- and Middle-Income Countries (LMICs) like India, where they constitute the majority of road traffic fatalities. A critical, yet under-explored, contributor to this risk is Time Pressure (TP)—the subjective perception of insufficient time to complete a task. TP induces cognitive stress, leading to risky behaviors such as overspeeding, abrupt braking, and reduced hazard response. While existing AI-driven traffic safety models exist, they primarily focus on four-wheelers or general traffic datasets. They often fail to model the unique, unstable dynamics of PTWs and do not adequately incorporate latent cognitive stressors like TP into collision risk assessment frameworks. Furthermore, many state-of-the-art deep learning models are computationally heavy, making them unsuitable for real-time, edge-deployable safety systems on low-cost hardware.

2. Methodology

2.1. Data Collection and Dataset

The authors constructed a large-scale, labeled multivariate time-series dataset to address the lack of PTW-specific data under varying TP conditions.

Participants: 51 male riders (ages 18–42) with at least two years of riding experience, reflecting the demographic most affected by PTW fatalities in India.
Experimental Setup: A high-fidelity static PTW simulator (Honda motorcycle frame with motion system) was used to ensure safety while capturing realistic kinematic data.
Conditions: Participants completed three riding sessions under No Time Pressure (NTP), Low Time Pressure (LTP), and High Time Pressure (HTP). The HTP condition simulated urgent scenarios (e.g., exam deadlines) to induce cognitive stress.
Data Volume: 153 sessions yielding 129,209 overlapping time-series windows (960 ms each, 100 Hz sampling).
Features: Each window contains 64 features categorized into vehicle dynamics, control inputs, proximity (headway/tailway), behavioral violations, and time context.
Labels: Binary collision labels (1 = collision event within the window, 0 = otherwise).

2.2. Proposed Architecture: MotoSafety

The paper introduces MotoSafety, a novel deep learning architecture designed for edge deployment, grounded in the principle of Learned Temporal Importance (LTI).

Parallel Multi-Scale Dilated Convolutional Encoder: Three parallel branches with dilation factors $d=\{1, 2, 4\}$ extract local-to-medium temporal patterns from the input sequence.
Bi-LSTM Branch: A two-layer Bidirectional Long Short-Term Memory network captures long-range sequential dependencies.
Temporal Importance Pooling (TIP): This is the core novel contribution. Instead of using fixed pooling (e.g., mean or max) or computationally expensive self-attention ( $O(L^2)$ ), TIP employs a lightweight Multi-Layer Perceptron (MLP) to learn a content-aware scoring function. It assigns an importance weight to every timestep independently for both the CNN and BiLSTM branches, collapsing them into fixed-size vectors. This reduces downstream fusion complexity to $O(1)$ relative to sequence length, maintaining an overall $O(L)$ complexity.
Fusion and Refinement: The pooled vectors are concatenated and processed through:
- Self-Gated Fusion: To dynamically weight feature combinations.
- Squeeze-and-Excitation (SE) Block: For channel-wise feature recalibration.
- Multi-Head Attention (MHA): For high-order feature refinement.
Output: The architecture supports two tasks:
1. Classification: Same-window collision risk detection (binary).
2. Forecasting: Multi-horizon kinematic state prediction (up to 7.2 seconds ahead).

2.3. Training and Evaluation

Loss Functions: Focal Loss ( $\gamma=2$ ) for classification to handle class imbalance; Mean Squared Error (MSE) for forecasting.
Baselines: The model was benchmarked against ten state-of-the-art models, including Transformer variants (Informer, iTransformer, PatchTST), frequency-based models (TimesNet), and LLM-based time-series models (Time-LLM, LLM4TS).
Hardware: Training on NVIDIA T400; inference tested on low-cost CPU (Intel Core i7) to simulate edge deployment.

3. Key Contributions

Large-Scale PTW Simulator Dataset: The first dataset of its kind capturing 129,000+ labeled sequences from 51 participants under No, Low, and High Time Pressure conditions, featuring 64 domain-specific features.
MotoSafety Architecture: A novel edge-AI framework integrating parallel dilated convolutions, BiLSTM, and the Learned Temporal Importance (TIP) module. TIP enables content-aware temporal collapse with linear complexity ( $O(L)$ ), making it suitable for real-time edge inference.
Performance and Efficiency: MotoSafety achieves 94.97% accuracy and 99.33% ROC AUC for collision risk assessment, outperforming ten baselines. For forecasting, it achieves 0.039 MSE and 0.094 MAE (4.4× lower error than Time-LLM and iTransformer).
Edge-Deployability: With only 1.15 million parameters and an inference latency of 0.135 ms on a CPU, the model is 21.9× faster than TimesNet and 71.9× faster than LLM4TS, validating its suitability for low-cost hardware.
Transferability: The architecture demonstrates strong generalization when retrained from scratch on human activity recognition (97.66% accuracy) and clinical exercise assessment (99.65% accuracy).

4. Results and Analysis

Impact of Time Pressure: Analysis confirms that High Time Pressure (HTP) significantly degrades rider behavior, resulting in 48% higher mean speeds, 36% more sudden braking, and 14% shorter headway distances compared to NTP.
TP as an Inductive Bias: Incorporating TP as a feature improves collision prediction accuracy from 94.09% (without TP) to 94.97% (with ground truth TP). Even when using a predicted TP label (derived from a separate Time Series Transformer), accuracy reaches 94.82%, demonstrating the value of modeling cognitive stress.
Real-World Feasibility: Using only 21 features available via standard low-cost IMU and GPS sensors (e.g., speed, acceleration, yaw rate), MotoSafety maintains 93.91% accuracy, suggesting practical deployment is feasible without high-fidelity simulator data.
Ablation Studies: Removing the TIP module and replacing it with standard pooling (mean, max, last-timestep) results in a performance drop of 1.56% to 2.74%, confirming that the performance gain stems from the learned, content-aware weighting rather than the architecture alone.

5. Significance and Claims

The paper claims that MotoSafety represents a significant advancement in Intelligent Transportation Systems (ITS) for vulnerable road users. By explicitly modeling the interplay between cognitive stress (Time Pressure) and operational maneuvers, the system moves beyond reactive crash detection to proactive risk assessment.

The authors emphasize that the model's lightweight design and low latency are critical for the "Safe System Approach," enabling the deployment of real-time alert systems (e.g., haptic or audio alerts on handlebars or smart helmets) in resource-constrained environments like LMICs. The work bridges the gap between theoretical deep learning and practitioner-validated solutions, offering a scalable framework to reduce PTW fatalities.

6. Limitations and Future Work

The authors acknowledge several limitations:

Demographics: The study is limited to male participants due to the statistical dominance of male PTW fatalities in India and the scarcity of female riders with manual transmission experience.
Environment: Data was collected in a simulator; while validated for behavioral patterns, real-world factors like weather and fatigue were not present.
Physiology: The study relies on behavioral markers for stress; physiological indicators (e.g., heart rate) were not measured.
Deployment: While the model is edge-ready, physical hardware validation on real motorcycles in traffic is identified as future work.

Future directions include expanding the cohort to include female riders, integrating physiological sensors, and conducting phased real-world trials on commercial riding platforms.