From Detection to Reliability: Supervised Contrastive Learning and Perception Evaluation in Dynamic Conditions for Autonomous Driving

Abstract

Safe autonomous driving requires both reliably detecting rare safety-critical events (SCEs) — crashes and near-crashes — and understanding when the perception systems responsible for detection can be trusted. This dissertation addresses this progression from detection to reliability. On the detection side, this work develops representation learning frameworks grounded in Supervised Contrastive Learning (SCL). Applied to the Second Strategic Highway Research Program (SHRP 2) Naturalistic Driving Study (NDS) video data, SCL enhances intra-class cohesion and inter-class separation in the learned representation space, improving crash/near-crash/baseline classification over supervised and self-supervised contrastive benchmarks. Extending this principle to vehicle kinematics, the proposed supervised contrastive variational autoencoder (scVAE) uses a dual-encoder architecture to disentangle a label-regulated salient latent space from shared driving patterns, achieving state-of-the-art clustering and enabling scenario generation, denoising, and event prediction. On the reliability side, this dissertation introduces Perception Characteristics Distance (PCD), a metric that quantifies the maximum distance at which a perception system remains reliable under a given decision rule. PCD models the mean and variance of perception performance as a function of distance via spline regression and variance change-point detection. Evaluated on the newly introduced SensorRainFall dataset across weather and lighting conditions, PCD reveals distance-dependent reliability characteristics that conventional static metrics overlook. Together, these contributions provide a unified statistical paradigm for both detecting SCEs and assessing perception reliability in dynamic conditions, supporting the development and validation of safe automated driving systems.