Vehicle-Level Temporal Image Sequences for Driving Behaviour Recognition

This dataset supports the study “Video-Based Detection of Risky Driving Manoeuvres from Vehicle-Level Temporal Sequences”. It contains 434 vehicle-level temporal image sequences representing both normal driving and risky driving behaviours, organised into 94 internally identified source-video groups. Each sequence corresponds to one tracked vehicle. The dataset was constructed through manual road-video selection, temporal trimming, video standardisation, vehicle detection, multi-object tracking, vehicle-sequence extraction, behavioural annotation, and temporal normalisation. YOLOv11 was used for vehicle detection, and ByteTrack was used to associate detections across consecutive frames. Sequences containing exactly 20 frames were retained, longer sequences were reduced through uniformly spaced frame selection, and shorter sequences were extended through interpolation based on adjacent-frame averaging. The original five-class label space comprises normal driving (351 sequences), brake check (53), cut off (21), hard braking (5), and zigzag (4). For the grouped three-class experiments, cut off, hard braking, and zigzag were combined into an OTHERS class. The dataset includes the vehicle-level image sequences and a semicolon-delimited metadata file, dataset_normalizado.csv, which provides the correspondence between each internal source-video identifier, tracked-vehicle identifier, and vehicle-level behavioural label. In the accompanying experiments, each image was resized to 160 × 160 RGB pixels and processed using a CNN-LSTM classifier that combines MobileNetV2 frame-level visual descriptors with bidirectional LSTM temporal modelling. Further information about provenance, copyright, licence scope, dataset structure, and limitations is provided in the included README.md file.