V-SenseDrive-PK: A Multimodal Benchmark Dataset for Driver Behaviour Analysis in Pakistan’s Urban and Highway Traffic

Abstract

Public naturalistic driving datasets come almost entirely from instrumented vehicles in high income countries, so no matching data exist for the mixed traffic and low cost sensing of low and middle income countries (LMICs). We release V-SenseDrive-PK , a smartphone dataset recorded on an urban secondary road and a highway in Islamabad, Pakistan. One phone logged inertial signals at 10 Hz and GPS at 1 Hz, and a second phone recorded 1080p road facing video. Four drivers in four vehicles each enacted normal, aggressive and drowsy sessions over 47 trips: 11.0 sensor hours, 11.7 video hours and 516 km. Video to sensor alignment is measured per video instead of assumed; after drift correction the median residual is 166 ms. The released rule based labels (Normal, Aggressive, Risky; drivers D1 to D3) turn out to be speed bands. A single 70 km/h rule produces 97% of Aggressive samples, speed separates the classes with \(\varepsilon ^2 = 0.54\) , and no acceleration or yaw channel exceeds \(\varepsilon ^2 = 0.001\) . Under leave one driver out evaluation XGBoost reaches macro F1 0.78 on the rule labels with speed and 0.36 without it, and 0.39 and 0.34 on the enacted session labels. With the classifier fixed, ResNet 18 frame embeddings do not improve on the sensor features. We release the labelled data and the per video synchronisation, with the raw logs, videos and evaluation code available on request, so that later work can separate speed as a risk factor from manoeuvre style, a distinction the current labels cannot make.

The authors' abstract, as published at the source. International Journal of Computational Intelligence Systems, 2026 · DOI ↗