Trial-Efficient Crash and Fall Detection for Two-Wheeler Delivery Riders Using Smartphone Inertial Sensors

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.