Jul 2026· International Journal for Research in Applied Science and Engineering Technology· Vol 14, pp. 2255-2263· 0 citations
TL;DR
This paper presents a machine learning pipeline for real-time crash and fall detection using only smartphone accelerometer and gyroscope data, without dedicated hardware, and proposes a tiered escalation architecture to absorb the false-positive cost while preserving high sensitivity to genuine falls.
Abstract
Delivery riders operating two-wheelers face a disproportionately high risk of road accidents, yet automatic crash
detection remains rare among Indian food and grocery delivery platforms. This paper presents a machine learning pipeline for
real-time crash and fall detection using only smartphone accelerometer and gyroscope data, without dedicated hardware. We
combine two public sensor datasets — a motorcycle-fall dataset collected via an instrumented motorcycle with staged real falls,
and a smartphone-based driver-behavior dataset — into a unified corpus of 21 independent trials and 1,788 sliding-window
feature vectors spanning four classes: normal riding, hard braking, pothole impact, and crash/fall. A Random Forest classifier,
selected over a comparably-performing XGBoost model for its interpretability, is evaluated using trial-grouped 5-fold crossvalidation to prevent leakage from overlapping sliding windows. The model achieves a mean cross-validated accuracy of 65.7%
(with high inter-fold variance attributable to limited trial diversity in two minority classes) and, more critically for the target
application, a mean recall of 92.3% on the crash/fall class with a false-positive rate of 14.8%. We propose a tiered escalation
architecture — an on-device check-in prompt followed by automatic emergency notification — to absorb the false-positive cost
while preserving high sensitivity to genuine falls. We report our findings transparently, including data-quality issues discovered
during preprocessing and the specific data-scarcity limitations driving evaluation variance, and outline a human-in-the-loop
retraining strategy as the direct path to improvement.
This study develops and validates a smartphone-based framework for automatically detecting emergency maneuvers, strong jolts, and crashes involving electric scooters and electric bicycles. Detection criteria were established through controlled track experiments and subsequently evaluated using data collected during a naturalistic riding study involving 119 participants and more than 26,000 km and 1,600 h of riding, combining accelerometer, gyroscope, GPS, and video recordings. Threshold-based detection criteria were defined using variables selected for their physical relevance and ability to discriminate between target and non-target situations. Hard braking, sharp turns, strong jolts, and crash-related events were identified using combinations of acceleration, jerk, rotational dynamics, and post-event vehicle motion. Video review showed that 74% of hard-braking detections corresponded to harsh-braking maneuvers, 64% of sharp-turn detections reflected genuine avoidance maneuvers, and 91% of strong-jolt detections were associated with infrastructure features. Video verification of collision candidates confirmed several reported and previously unreported impacts, including collisions with other road users and single-vehicle falls. Application of the framework to the naturalistic dataset revealed marked differences between vehicle types. E-scooter users experienced higher rates of hard braking and strong jolts than e-bicycle users, reflecting behavioral differences and vehicle characteristics. Illustrative mapping examples showed that detected events and rider-reported hazardous situations could occur in close proximity, suggesting opportunities for future spatial analyses of micromobility safety. Although additional validation on larger crash datasets is required, the results demonstrate that threshold-based approaches can provide meaningful indicators of rider safety, support large-scale monitoring of micromobility risks, and contribute to infrastructure and transport-safety assessment.
C. Naude, Ebrahim Riahi, Bastien Canu et al.· Accident Analysis and Preven...· 0 citations
These findings demonstrate that RF–Bayesian provides a stable, interpretable, and computationally efficient framework for smartphone-based driver behavior classification, with practical relevance for telematics, fleet safety management, driver feedback systems, and intelligent transportation safety applications.
A.A. Al-Rababah, S. M. Rahman· Neural computing & applicati...· 0 citations
Cycling is increasingly promoted as part of sustainable transport strategies, but safety concerns remain an important barrier. Crash-based safety assessment is limited because crashes are reactive indicators and cycling crashes are often underreported. This study evaluates whether smartphone inertial sensor data can be used to detect selected potentially critical cycling manoeuvres that may support proactive safety screening. A two-stage workflow intended for possible future smartphone implementation was evaluated offline. In Stage 1, kinematic thresholds were applied to identify short candidate sensor windows when a high-dynamic manoeuvre was detected. In Stage 2, a machine learning classifier was applied to these candidate windows to classify the manoeuvre type and reduce false positives. The workflow was evaluated using three datasets collected in Germany from 10 riders, including controlled test-track manoeuvres and urban rides in Aachen and Bonn. The dataset included 148 hard-braking and 132 abrupt-swerving manoeuvres. In addition, 45 bicycle-only crash-like reference sequences were collected to compare low-speed impact-like signal patterns. In leave-one-dataset-out evaluation, the Support Vector Machine achieved mean precision–recall areas under the curve of 0.954 for hard braking and 0.884 for abrupt swerving. The corresponding mean F1-scores were 0.924 and 0.876. The offline evaluation showed that smartphone sensor data can support the detection of selected high-dynamic cycling manoeuvres with good classification performance under controlled and semi-controlled conditions. However, most target manoeuvres were controlled or rider-initiated proxy manoeuvres rather than confirmed traffic conflicts. The results should therefore be interpreted as manoeuvre detection, not as direct crash-risk validation. In future naturalistic studies, the workflow could help identify candidate locations where cyclists frequently brake hard or swerve abruptly. These locations can then be prioritised for site inspections, video-based conflict analysis, or comparison with crash records.
Jörg Ehlers, A. García-Hernandez, A. Wolter· Italian National Conference...· 0 citations
This study analyses 397,850 police-reported crashes in Sri Lanka from 2010 to 2020 using a multinomial logit model, validated through Random Forest and XGBoost, to identify key factors influencing crash severity across four outcome levels. Results reveal four findings not apparent from conventional crash statistics. First, motorcycle–vulnerable road user crashes (OR = 202, 10.24% of crashes) produced higher fatal odds than heavy vehicle–motorcycle interactions (OR = 143, 6.86%), despite the latter receiving greater policy attention. This suggests that motorcycle-related risk to vulnerable road users is considerably underappreciated in current safety programs. Second, single-vehicle motorcycle crashes (OR = 46.2, 9.05%) were three times more fatal per incident than motorcycle–passenger vehicle collisions (OR = 14.6), implicating loss of control and emergency response delays as under-recognized contributors. Third, passengers falling from buses yielded a fatal odds ratio of 2,337, the highest in the model, despite representing only 1.50% of crashes, indicating a concentrated fatality risk understated in national reporting. Fourth, dusk and dawn conditions (OR = 2.24, 17.27% of crashes) were found to be equally as dangerous as completely unlit nighttime roads (OR = 2.26), while fog and mist produced the highest lighting-related fatal odds ratio of 4.71. The findings indicate that targeted interventions addressing motorcycle–vulnerable road user interactions, bus passenger safety, and transitional lighting conditions offer the greatest potential for reducing fatal crash outcomes in Sri Lanka.
D. Bhagya, N. Jayantha, H. Pasindu· Proceedings of the 19th Tran...· 0 citations
This research establishes a statistically robust and deployable foundation for next-generation intelligent transportation systems by coupling Bayesian learning theory with edge computing design.
Seyed Mohammad Hosseini, V. Kiani, Hadi Sadoghi-Yazdi· Computing· 0 citations
Drowsiness is a leading cause of human error in transportation and in shift-based occupational work, yet delivering reliable real-time detection on affordable, resource-constrained hardware remains difficult. This study aims to develop and evaluate a vision-based drowsiness detection system that behaves consistently across a full-power desktop and a low-cost edge device. The system couples two independent MobileNetV2 transfer-learning classifiers — one for eye state (Open/Closed) and one for mouth state (Yawn/No_yawn) — with a temporal decision engine that converts frame-level predictions into microsleep and excessive-yawning alerts. Both classifiers were trained on a merged multi-source dataset (8,548 training images) and evaluated with a class-balanced protocol (186 images/class for the eye branch and 448 images/class for the mouth branch) to remove test-set imbalance bias. The decision engine was realised as two platform-appropriate pipelines that share an offline-first, retry-capable event architecture: a duration-based, two-tier hysteresis alert on a desktop application (Haar-cascade detection, H5/float32 models) and a frame-count alert designed for a Raspberry Pi 5 edge board (MediaPipe detection, quantised TensorFlow Lite models). On the class-balanced test set the eye branch reached 95.16% (H5) / 95.97% (TFLite) accuracy and the mouth branch reached 96.65% for both formats, with above-99% cross-format prediction agreement. Converting to TFLite cut model size by 73.4% (8.99 to 2.39 MB) and single-frame model inference latency roughly thirteen-fold (about 25 to 1.9 ms, measured on the development machine). Real-time desktop sessions sustained 6.9–12.0 FPS, and the Haar detector located a face in only 15.8% of off-angle frames versus 96.9–98.6% of frontal frames. This single-session result offers a preliminary, rather than definitive, indication of the detector's pose sensitivity. A lightweight dual-MobileNetV2 design with platform-appropriate detectors shows promise for delivering consistent real-time drowsiness alerts across heterogeneous hardware tiers.
Rafi'e· Indonesian Journal of Electr...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
