Toward Transparent Perception: An Explainable Artificial Intelligence Framework for Object Detection in Autonomous Vehicles

Abstract

The rapid integration of deep learning based object detection into autonomous vehicle perception systems has improved
detection accuracy substantially, yet the opacity of these models continues to constrain their acceptance among regulators,
engineers, and the travelling public. This paper undertakes a theoretical and conceptual synthesis of the expanding literature
on explainable artificial intelligence as it applies specifically to object detection within autonomous driving contexts. Drawing
on a structured narrative review of peer-reviewed scholarship spanning computer vision, human-computer interaction, and
artificial intelligence governance, the study examines how post-hoc explanation techniques, attention-based architectures, and
counterfactual reasoning methods have been adapted to the particular demands of real-time vehicular perception. The review
identifies persistent tension between explanatory fidelity and computational latency, inconsistent standards for evaluating
explanation quality, and a scarcity of frameworks that connect technical explainability to the trust calibration processes
of human operators and pedestrians. Building on these findings, the paper proposes a layered conceptual framework that
situates explanation generation as an intermediary function between perception and human-machine interaction rather than
as an auxiliary diagnostic feature appended after deployment. The framework is illustrated through a comparative analysis
of five widely referenced explainability techniques evaluated against interpretability, latency, and deployment suitability.
The paper concludes that explainability in autonomous vehicle object detection should be treated as a design requirement
embedded from the earliest stages of system architecture, carrying implications for regulatory certification, public trust, and
the direction of future empirical research.