Answers are grounded in this paper and cite their sources.
Object detection is the safety-critical core of autonomous vehicle (AV) perception, yet benchmark accuracy alone provides limited evidence of real-world deployment readiness. A detector that performs well on public datasets may still fail under adverse weather, miss distant pedestrians, or exceed the latency constraints of embedded automotive hardware. Although existing reviews comprehensively describe the evolution of CNN, Transformer, and hybrid-based detectors, they do not systematically evaluate deployment evidence or account for inconsistencies in datasets, hardware platforms, and evaluation protocols. This review addresses these limitations through three original contributions. First, a unified taxonomy categorizes detector families according to the integration of local convolutional and global attention mechanisms and relates these design choices to AV-specific operational challenges. Second, an evidence-comparability protocol defines the conditions under which reported accuracy, latency, and robustness can be validly compared, exposing common sources of misleading cross-study comparisons. Third, the proposed AV Object Detection Deployment-Readiness Framework (AVOD-DRF) transforms published evidence into an eleven-dimension deployment-readiness profile, where each score is explicitly evidence-linked and missing evidence is distinguished from poor performance. These contributions are supported by a PRISMA 2020-guided systematic review of six bibliographic databases covering 2015 to 2026, resulting in 142 eligible studies. The analysis reveals substantial reporting gaps: only 13% evaluate latency on automotive or edge hardware, 19% assess adverse weather or low-light conditions, and 15% perform cross-dataset validation. No detector family is universally superior; rather, deployment suitability depends on application constraints. The proposed framework provides a reproducible, evidence-based methodology for selecting AV object detectors based on real-world deployment readiness rather than benchmark performance alone.
No takes yet. Share an insight, caveat, or question.
Chiddarwar et al. (2026) studied this question.