cs.CVAug 19, 2026
SaveOne-Stage Object Detectors in Autonomous Driving
Organizations: Florida Gulf Coast University
Abstract
Autonomous vehicles depend on fast and reliable perception systems to detect surrounding vehicles, pedestrians, cyclists, traffic signs, and other road objects in real time. This paper presents a comprehensive survey and analysis of one-stage object detectors for autonomous driving rather than an implementation of a new detection system. The survey reviews the evolution of major one-stage detectors, including YOLOv1, SSD, RetinaNet, EfficientDet, anchor-free detectors such as FCOS and CenterNet, and recent real-time models such as YOLOv10. It compares these architectures through their design choices, feature-fusion strategies, loss functions, deployment trade-offs, and reported benchmark performance. The paper also summarizes commonly used autonomous-driving datasets, evaluation metrics, open challenges, and future research directions. Overall, this survey highlights how one-stage detectors balance speed, accuracy, efficiency, and robustness, while also emphasizing the remaining gap between benchmark results and dependable real-world autonomous-driving performance.
Figures & tables
| Paradigm | Architectural Family | Sub-Family Classification | Representative Architectures |
| Two-Stage Detectors | R-CNN Series | Region Proposal | Faster R-CNN, Mask R-CNN, Cascade R-CNN |
| One-Stage Detectors | Anchor-Based | Default Box Matching | SSD, RetinaNet, EfficientDet [ 33 , 39 , 7 ] |
| Grid-Based Anchors | YOLOv2, YOLOv3, YOLOv4, YOLOv5, YOLOv6, YOLOv7 [ 27 , 28 , 29 , 31 ] | ||
| Anchor-Free | Center-Based | YOLOv1, FCOS, YOLOX, YOLOv8, YOLOv9, YOLOv10 [ 26 , 38 , 30 , 32 ] | |
| Keypoint-Based | CornerNet, CenterNet, ExtremeNet [ 36 , 37 ] |
TABLE I: Taxonomy of Object Detection Methods in Autonomous Driving
| Sub-Family | Core Design Principle | Key Detectors That Belong Here |
| Anchor-Free / Point-Based | Eliminates predefined anchor boxes to reduce complexity and improve localization of irregular objects | FCOS (ReFPN-FCOS) |
| Efficiency-Optimized | Focuses on reducing parameter counts and FLOPs for deployment on edge hardware. | EfficientDet, ShuffYOLOX |
| NMS-Free / Real-Time YOLO | Utilizes dual-assignment strategies to remove inference bottlenecks and reach ultra-low latency. | YOLOv10, YOLOX |
TABLE II: Three sub-families of one-stage detectors
Fig. 1: Timeline of one-stage detectors from 2016 to 2024. The figure is placed across both columns to improve readability. If the source graphic is regenerated, use larger labels and more spacing between model boxes.
| # | Detector | Year | Sub-Family | Key Innovation |
| 1 | YOLOv1 | 2016 | YOLO Family | Increasing the real-time speed and detection accuracy |
| 2 | YOLOv2 | 2017 | YOLO Family | Better accuracy |
| 3 | YOLOv3 | 2018 | YOLO Family | Multi-scale prediction |
| 4 | YOLOv4 | 2020 | YOLO Family | An increase in detection accuracy |
| 5 | YOLOX | 2021 | YOLO Family | An anchor-free formulation and SimOTA label assignment |
| 6 | YOLOv7 | 2023 | YOLO Family | Architectural refinements |
TABLE III: Summary of the selected one-stage detectors
| Dataset | Year | # Images | # Classes | Key AV Focus / Characteristics | Commonly Used Metric |
| KITTI | 2012 | 7,481 train / 7,518 test | 8 | Urban driving benchmark with cars, pedestrians, cyclists, vans, trucks, trams, and other road objects. Often used for 2D/3D detection, depth, tracking, and distance estimation. | AP, mAP, 3D AP |
| Waymo Open Dataset | 2019 | 1M+ camera images | 4 main detection classes | Large-scale real-world driving dataset with camera and LiDAR data across diverse road scenes. Useful for evaluating detection in complex traffic environments. | mAP, mAPH |
| nuScenes | 2019 | 1.4M camera images | 10 detection classes | Multi-sensor dataset with cameras, LiDAR, radar, GPS, and IMU. Strong for autonomous driving because it evaluates detection across full 360-degree driving scenes. | mAP, NDS |
| BDD100K | 2020 | 100,000 images | 10 detection classes | Large driving dataset with diverse weather, lighting, road types, and time-of-day conditions. Useful for studying real-world robustness. | mAP |
| Cityscapes | 2016 | 5,000 finely annotated images | 8 instance-level classes | Urban street-scene dataset focused on dense pixel-level understanding. Often used for segmentation, but still useful for evaluating road-scene perception quality. | mIoU, AP |
| Argoverse | 2019 | 290,000+ images | 15 | Autonomous driving dataset focused on 3D tracking, maps, and urban driving scenes. Useful for detection and motion understanding in city environments. | mAP, tracking metrics |
TABLE IV: Benchmark datasets used in autonomous driving object detection
Fig. 2: Comparison of major benchmark datasets used for autonomous driving detection.
| Metric | Full Name | What It Measures | Higher is Better? |
| mAP | mean Average Precision | Shows how well the detector finds objects and labels them correctly across all classes. | Yes |
| AP | Average Precision | Measures how well the detector finds one specific object class, such as cars or pedestrians. | Yes |
| IoU | Intersection over Union | Checks how much the predicted box overlaps with the real object box. | Yes |
| FPS | Frames Per Second | Shows how many images the model can process each second. | Yes |
| Latency | Inference Delay | Measures how long the model takes to make one prediction. | No |
| Parameters | Model Parameters | Shows how large the model is based on how many learned values it has. | No |
TABLE V: Evaluation metrics used in this survey
| Detector | Year | Backbone | Dataset | mAP/AP (%) | FPS | Params (M) | Input Size |
| YOLOv1 [ 26 ] | 2016 | Custom CNN | VOC 2007 | 63.4 | 45 | N/R | 448 448 |
| YOLOv2 [ 27 ] | 2017 | Darknet-19 | VOC 2007 | 76.8 | 67 | N/R | 416 416 |
| YOLOv3 [ 28 ] | 2018 | Darknet-53 | COCO | 33.0 | 20 | N/R | 608 608 |
| YOLOv4 [ 29 ] | 2020 | CSPDarknet53 | COCO | 43.5 | 62 | 64 | 608 608 |
| YOLOX [ 30 ] | 2021 | CSPDarknet | COCO | 50.0 | 68.9 | 54 | 640 640 |
| YOLOv7 [ 31 ] | 2023 | E-ELAN | COCO | 51.4 | 161 | 36.9 | 640 640 |
TABLE VI: Performance comparison of one-stage detectors. Values are reported from the original papers when available. Because the papers use different datasets, metrics, input sizes, and hardware, the results should be interpreted as reported benchmark values rather than a perfectly controlled comparison.
Fig. 3: Speed-accuracy trade-off plot of one-stage detectors.
| Detector | Key Advantages | Key Disadvantages | AV Deployment Limitation |
| YOLO (v4 / v8 / v10 family) | Very high inference speed, making it real time capable. single-stage architecture. strong speed and accuracy balance. scalable (small can turn into large variants). efficient for edge deployment. | Slightly lower localization accuracy than two-stage models. struggles with very small or occluded objects. more performance sensitive to dataset quality than other models. | May miss small/distant objects (e.g., pedestrians, bikes) in dense traffic. requires careful tuning for safety critical scenarios, such as unique problem cases. |
| SSD (Single Shot Detector) | Fast inference (often the highest FPS of all one-stage models). Lightweight. Simple architecture. suitable for embedded/low-power devices. More efficient for computers. | Lower accuracy compared to YOLO and EfficientDet. weak performance on small objects. limited feature representation depth. | High miss rate for small/critical objects which reduces reliability in AV perception pipelines. |
| RetinaNet | Good balance between accuracy and handling class imbalance (Focal Loss). strong detection of hard/rare objects. | Slower than single-stage models like YOLO and SSD. computationally heavier. still struggles with latency constraints. | Difficult to meet real-time AV requirements without optimization. limited edge deployment feasibility. |
| RefineDet | Improves SSD with two-step refinement. better localization accuracy than SSD. maintains relatively high speed. Multi-stage refinement improves detection quality. | More complex than SSD. slower than pure one-stage models. still weaker than YOLO in speed; increased training complexity. | May not meet strict real-time latency for high-resolution AV inputs. harder to deploy on low-power edge hardware. |
| CornerNet | Anchor-free detection (predicts object corners). Reduces dependence on anchor tuning. Good localization for irregular shapes. | Computationally expensive. slow inference. Grouping corners is complex. sensitive to detection errors at corners. | Not suitable for real-time AV systems. latency too high for onboard perception pipelines. |
| CenterNet | Detect objects via center points. simpler than corner-based methods. A good balance between accuracy and speed. fewer hyperparameters (no anchors). | Still slower than YOLO. struggles with dense object clustering. center-point ambiguity in crowded scenes. | May miss closely packed objects (e.g., urban traffic). borderline real-time without optimization. |
TABLE VII: Advantages, disadvantages, and deployment limitations of one-stage detectors
Fig. 4: Radar chart comparing representative one-stage detectors across multiple dimensions.
| Challenge | Description | Why It Matters for AVs | Partially Addressed By |
| Speed vs. Accuracy Trade-off | Faster models sacrifice precision, especially for small or complex objects | Missed or incorrect detections can lead to unsafe driving decisions | YOLO (balanced), EfficientDet (scaling trade-off) |
| Small Object Detection & Occlusion | Difficulty detecting small, distant, or partially hidden objects | Critical objects like pedestrians or bikes may be missed | YOLO (multi-scale), FCOS (feature pyramids), CenterNet |
| Robustness to Environmental Conditions | Performance drops under poor lighting, weather, and sensor noise | AVs must operate reliably in all conditions for safety | YOLOv5 (robustness improvements), data augmentation methods |
| Resource-Constrained Deployment | Limited compute, memory, and power on edge devices | Prevents real-time inference or requires heavy optimization | SSD (lightweight), YOLOv5/YOLOv8 (efficient variants) |
| Lack of Deployment-Centric Metrics | Over-reliance on mAP and FPS ignores real-world constraints | Misleading evaluation of model readiness for AV systems | EfficientDet (efficiency-aware design), recent benchmarking studies |
| Anchor-Free Localization Limitations | Anchor-free models struggle in dense or overlapping scenes | Incorrect localization affects tracking and decision-making | FCOS, CenterNet (partial solutions) |
TABLE VIII: Open challenges in one-stage detection for autonomous driving
More in this topic
Object Detection
Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.
See all 624 papers