Vision-Based Detection of Reckless E-Scooter Riding

Sammanfattning

Detecting reckless e-scooter riding from naturalistic front-facing camera video is challenging due to the scarcity of positive examples and the visual diversity of reckless behaviours. This thesis evaluates and compares unsupervised and supervised approaches to reckless riding detection, all based on a shared frozen video representation derived from a pre-trained video masked autoencoder. The thesis systematically compares scoring strategies under severe positive-label scarcity, examining how temporal granularity and the availability of labelled data affect detection performance. Methods are evaluated on a dataset of 458 rides, of which six are reckless, using average precision over 30 repeated evaluations. Conformal prediction is further applied to convert continuous ride scores into binary decisions at a controlled false positive rate.The results indicate that the frozen representation supports detection without domain-specific adaptation. Unsupervised anomaly scoring achieves strong performance relative to the baselines without labelled data during training, reaching a mean average precision of 0.519. Supervised methods reach a mean average precision of up to 0.671 but exhibit considerably higher variance across seeds. Conformal thresholding produces empirical false positive rates close to the nominal levels across all methods.