Technical Summary: A PRISMA-Guided Critical Review of CNN, Transformer, and Hybrid Object Detection Architectures for Autonomous Driving

Problem Statement

Object detection serves as the safety-critical perceptual foundation for autonomous vehicles (AVs), yet the field suffers from a disconnect between benchmark accuracy and real-world deployment readiness. While existing literature extensively catalogs the evolution of Convolutional Neural Networks (CNNs), Transformers, and hybrid architectures, these reviews often fail to systematically evaluate deployment evidence. A significant gap exists where detectors performing well on public datasets may fail under adverse weather, miss distant pedestrians, or exceed the latency constraints of embedded automotive hardware. Furthermore, cross-study comparisons are frequently invalid due to inconsistencies in datasets, hardware platforms, numerical precisions, and evaluation protocols. The core problem addressed is the lack of a systematic, evidence-based methodology to determine if a detector is truly ready for deployment, rather than merely ranking models based on benchmark metrics alone.

Methodology

This review employs a PRISMA 2020-guided systematic approach to ensure reproducibility and auditability.

• Search Protocol: Six bibliographic databases (Scopus, Web of Science, IEEE Xplore, ACM Digital Library, ScienceDirect, SpringerLink) were searched from January 2015 to June 2026. The search targeted peer-reviewed articles and full conference papers focusing on object detection in autonomous driving scenarios.
• Selection Criteria: Studies were included if they evaluated object detection in road-traffic scenes, reported defined architectures with quantitative metrics, and identified datasets. Camera-LiDAR and 3D detection studies were retained only as contextual categories, with the core synthesis focusing on camera-based 2D detection.
• Data Synthesis: From an initial pool of 1,980 records, 142 eligible studies were included after deduplication and screening.
• Comparability Protocol: A strict evidence-comparability protocol was established. Results are only considered comparable if dataset splits, metric definitions, IoU thresholds, input resolutions, hardware platforms, numerical precisions, and inference runtimes match. Where these conditions are not met, results are synthesized as evidence ranges or qualitative statements rather than ranked numerically.
• Quality Assessment: A ten-criterion rubric was applied to score studies on dataset transparency, metric definition, hardware reporting, reproducibility, and robustness evaluation.

Key Contributions

The paper makes three primary contributions to the field:

1. Unified Taxonomy: A classification system that categorizes detector families not merely by labels (e.g., "YOLO" or "DETR") but by how and where they integrate local (convolutional) and global (attention) representations. This taxonomy links specific design choices (sequential, parallel, or interleaved integration) to AV-specific operational challenges and failure conditions.

2. Evidence-Comparability Protocol: A set of rules defining the conditions under which reported accuracy, latency, and robustness can be validly compared. This protocol exposes common sources of misleading cross-study comparisons, such as quoting desktop GPU frame rates as evidence of embedded feasibility.

3. AV Object Detection Deployment-Readiness Framework (AVOD-DRF): A novel methodology that transforms published evidence into an eleven-dimension deployment-readiness profile. These dimensions include:

   • Accuracy on driving data
   • Small-object sensitivity
   • Inference latency on named hardware
   • Hardware dependency (precision, runtime, batch size)
   • Parameter and memory load
   • Robustness to weather and illumination
   • Resistance to occlusion and domain shift
   • Cross-dataset generalization
   • Uncertainty reporting and calibration
   • Reproducibility (code, weights, configuration)

   Crucially, the framework distinguishes between "missing evidence" (scored as 0) and "poor performance," ensuring that benchmark-focused studies are not penalized for questions they did not pose.

Key Results and Findings

The analysis of the 142 studies reveals substantial reporting gaps and specific architectural trade-offs:

• Reporting Gaps:
  • Only 13% of studies evaluate latency on automotive or edge hardware.
  • Only 19% assess adverse weather or low-light conditions.
  • Only 15% perform cross-dataset validation.
  • Only 6% report uncertainty estimation.
  • Only 5% report energy or power consumption.
• Architectural Trade-offs:
  • CNNs (One-Stage): Dominate latency-sensitive deployments due to mature toolchains, deterministic runtime, and efficient embedded optimization (e.g., INT8 quantization). However, they struggle with global context, small objects at long ranges, and distribution shifts (adverse weather).
  • Transformers (Query-Based): Offer superior performance in crowded and occluded scenes due to global self-attention and set prediction (eliminating NMS). However, they face challenges with small-object accuracy (unless multi-scale attention is used), high memory consumption, and immature operator support on embedded hardware.
  • Hybrids: Combine local and global reasoning. While they improve the accuracy-compute frontier on benchmarks, their deployment costs (quantization difficulty, operator fusion issues) are systematically under-measured. No study independently replicates a proprietary hybrid AV detector, leaving industrial claims unverifiable.
• Robustness: All families degrade substantially under adverse conditions, domain shifts, and adversarial attacks. No single family is universally robust; "robustness" is often conflated across different failure sources (natural corruption vs. adversarial perturbation) in the literature.
• Real-Time Feasibility: Only mature one-stage CNNs have demonstrated real-time operation on named embedded automotive hardware with stated precision. Transformer and hybrid detectors have demonstrated desktop throughput, but their embedded readiness cannot yet be certified from published evidence.

Significance and Claims

The paper explicitly states that it does not propose a new detector. Instead, its significance lies in providing the synthesis, comparability rules, and decision logic that previous reviews lack.

• Shift in Focus: The review argues that progress toward deployable AV perception now depends less on further benchmark accuracy and more on standardized evidence (latency on named hardware, condition-controlled robustness, cross-dataset degradation, and calibrated uncertainty).
• Actionable Framework: The AVOD-DRF is offered as a reusable methodology for researchers to profile new detectors, for AV teams to compare candidates against Operational Design Domain (ODD) requirements, and for reviewers to audit deployment claims.
• Research Agenda: The paper concludes with a six-point research agenda paired with falsifiable validation protocols, calling for:
  1. Standardized adverse-condition benchmarking.
  2. Hardware-normalized latency and energy reporting.
  3. Cross-dataset and cross-geography validation.
  4. Open-world and vision-language detection for unknown hazards.
  5. Uncertainty-aware and certifiable detection.
  6. Continual and fleet learning with forgetting safeguards.

The authors conclude that while no detector family is universally superior, deployment suitability is strictly dependent on application constraints. The proposed framework allows for an evidence-based selection of AV object detectors based on real-world readiness rather than benchmark performance alone.