1. Introduction

The rapid development of autonomous driving technology enables vehicles to navigate safely in highly complex and dynamic environments. As a crucial component of the perception system, object detection is responsible for real-time monitoring and analysis of various objects on the road, ensuring safe driving. According to existing research, the accuracy of object detection algorithms directly affects the safety and reliability of autonomous driving systems. In recent years, the continuous advancement of deep learning technologies has led to the emergence of many novel object detection algorithms, significantly enhancing detection performance.

RetinaNet, proposed by Lin et al., addresses the common issue of class imbalance in object detection by introducing the concept of Focal Loss. This characteristic is particularly important in real-world scenarios since the frequency of different object classes often varies significantly. This capability allows RetinaNet to effectively improve the recognition of low-frequency targets, thereby enhancing overall detection performance [1]. PointPillars, developed by Lang et al., focuses on the efficient processing of LiDAR point cloud data. By converting point cloud data into a pillar format, this algorithm significantly accelerates 3D object detection. The study demonstrates that, in autonomous driving environments, both detection accuracy and real-time processing capacity are critical [2]. Therefore, PointPillars achieves new heights in speed while effectively enhancing target detection accuracy in complex environments. Subsequently, CenterNet, proposed by Zhou et al., serves as an innovative object detection method based on center point detection, emphasizing effective recognition and detection of small objects in dense scenes. CenterNet transforms object detection into center point prediction, allowing accurate localization and identification of multiple targets in complex autonomous driving environments, which is essential for real-time system response [3]. Furthermore, the You Only Look Once (YOLO)algorithm introduced by Chen et al. achieves a good balance between speed and accuracy through novel structural designs and training strategies. This method has been optimized for real-time object detection scenarios, becoming a commonly used solution in autonomous driving visual systems [4].

Although existing object detection algorithms have made progress in accuracy and speed, they still face challenges in complex driving environments, such as changes in lighting, occlusion effects, and real-time processing requirements. Therefore, this paper will delve into the analysis of three representative object detection algorithms: YOLO, RetinaNet, and CenterNet. It will discuss their workflows, characteristics, and practical application performance in autonomous driving. The article will first outline the basic frameworks of these three algorithms, followed by discussions on their actual usage and improvements in the autonomous driving domain. Finally, experimental results will be presented to illustrate the performance of these algorithms in practical applications, revealing the current research status of object detection in autonomous driving and future development directions.

2. Overview of Basic Algorithms

2.1. You Only Look Once

YOLO algorithm transforms the object detection problem into a regression problem, aiming to achieve real-time object detection. Traditional object detection methods typically require generating candidate regions, followed by classification and regression for each region, which incurs a speed bottleneck. YOLO resolves this issue by dividing the entire image into an SxS grid and directly predicting bounding boxes and class probabilities within each grid. Its core task is to facilitate fast and accurate object detection, especially in applications requiring real-time feedback, such as video surveillance and autonomous driving.

The workflow of YOLO includes several key steps. First, the input image is divided into an SxS grid, with each grid responsible for predicting objects located within it. For each grid, YOLO simultaneously predicts B bounding boxes and their confidence scores (indicating the probability that an object exists within the box) and class probabilities. The algorithm then computes each bounding box's coordinates and dimensions and applies Non-Maximum Suppression (NMS) to eliminate overlapping boxes, retaining only the candidates with higher confidence scores. By this means, YOLO accomplishes object detection in a single forward pass, significantly enhancing processing speed. The advantages of YOLO lie in its rapid detection speed, meeting real-time processing requirements, and suitability for dynamic scenarios. However, YOLO's performance in small target detection is relatively weak, and because it uses the entire image as input, it may lead to insufficient understanding of complex scenes. Additionally, its accuracy may not always match that of two-stage algorithms.

2.2. RetinaNet

RetinaNet is designed to address the class imbalance problem in single-stage object detection algorithms, especially in small target detection, where the ratio of positive to negative samples is often severely imbalanced. To this end, RetinaNet introduces the Focal Loss function, which allows the network to focus more on hard-to-classify samples, thereby enhancing detection accuracy. The core task of this algorithm is to improve small target detection capability while maintaining high real-time performance.

The workflow of RetinaNet includes several critical steps. First, the algorithm employs a Feature Pyramid Network (FPN) for image feature extraction, combining high-level and low-level features to enable detection at different scales. Next, RetinaNet generates anchor boxes at each feature level and classifies and regresses each anchor box. The introduction of Focal Loss ensures that the network pays more attention to difficult samples during training, alleviating the impact of easy samples on training. Finally, RetinaNet also employs NMS to eliminate redundant boxes, ensuring that the final output detection results are optimal. The advantages of RetinaNet include a significant improvement in small target detection accuracy while maintaining high real-time performance. Through the introduction of Focal Loss, RetinaNet excels in addressing the class imbalance issue. However, compared to other single-stage algorithms, RetinaNet's response speed may be slightly slower, and its performance in extremely complex scenes still requires further optimization.

2.3. CenterNet

CenterNet is an advanced object detection algorithm that centers around detecting the center points of objects. Unlike traditional frameworks, CenterNet directly performs detection by regressing the center points of objects and their associated attributes. The method mainly consists of three steps: first, features are extracted through a backbone network; second, convolutional layers generate a center point heatmap and regression maps for other related attributes (such as bounding box width and height); finally, post-processing is conducted to obtain the final prediction results. CenterNet's key features include its strong capabilities in handling small objects and dense targets. Additionally, its relatively low computational overhead allows it to maintain high frame rates in real-time applications.

3. Applications and Improvements of Algorithms in Autonomous Driving

3.1. Applications and Improvements of YOLO in Autonomous Driving

The YOLO algorithm is widely used in the environmental perception of autonomous vehicles due to its efficiency and real-time capabilities. To enhance YOLO's performance in complex urban environments, researchers have implemented various improvements. For example, Alexey Bochkovskiy et al. proposed several optimization techniques, such as data augmentation, model feature fusion, and improved background modeling in their work "YOLOv4: Optimal Speed and Accuracy of Object Detection." These improvements aim to address YOLO's accuracy issues in dense target scenarios [5]. Specific improvements include:

3.1.1. Data Augmentation. Methods such as mirroring, rotation, and color adjustment expand the training dataset, improving the model's adaptability to new scenarios. Glenn Jocher et al. noted in "YOLOv5: Better, Faster, Stronger" that increasing data diversity and richness enhances the model's generalization ability and recognition rates in different environments [6].

3.1.2. Feature Fusion. Utilizing Feature Pyramid Networks (FPN) to combine multi-scale features enhances small object detection capabilities. For instance, in "Real-time Object Detection in Autonomous Driving Based on Improved YOLOv3 Algorithm," researchers improved YOLO's detection accuracy through feature fusion, particularly for small targets (e.g., pedestrians and traffic signs) [7].

3.1.3. Post-Processing Optimization. By refining the Non-Maximum Suppression (NMS) strategy, the number of overlapping detections is reduced, resulting in higher accuracy. The optimized NMS effectively reduces the false alarm rate in YOLO’s practical applications.

These improvements have enhanced YOLO's accuracy in complex scenarios. Experimental results show that the optimized YOLOv4 achieves an Average Precision (mAP) of 79.5% on the KITTI dataset, an approximate 4% improvement over YOLOv3. Additionally, YOLOv5 maintains a frame rate of over 35 FPS, making it suitable for real-time autonomous driving applications.

3.2. Applications and Improvements of RetinaNet in Autonomous Driving

The RetinaNet algorithm is particularly well-suited for handling class imbalance and small object detection issues. In the face of complex scenarios in autonomous driving, researchers have made numerous improvements to RetinaNet's feature extraction and loss functions.

3.2.1. Focal Loss Function Optimization. The core innovation of RetinaNet lies in its use of Focal Loss. This mechanism weights background regions that are often overlooked, significantly enhancing small object detection accuracy. Lin et al. elucidated in "RetinaNet: Focal Loss for Dense Object Detection" how this mechanism effectively addresses small and minority class targets [1].

3.2.2. FPN Enhancement. Improving the feature pyramid enables multi-scale information extraction, increasing the model's capability to recognize objects of various sizes in complex environments. In "An Efficient Real-Time Object Detection Algorithm Based on RetinaNet for Autonomous Driving," researchers enhanced small object detection capability in real-world scenarios through FPN improvements [8].

3.2.3. Data Augmentation and Loading Strategies. Incorporating diverse datasets and additional data augmentation techniques enhances the model’s robustness and accuracy in complex scenes. For instance, Zhang et al. reported in "Object Detection in Autonomous Driving Using RetinaNet and Improved Data Augmentation" that improved data loading strategies elevate training efficiency and final model performance [9].

Experimental results indicate that the enhanced RetinaNet achieves a 7% increase in Average Precision for small object detection on the KITTI dataset. Furthermore, the modified algorithm demonstrates significantly improved recognition accuracy during driving tests in complex urban environments, ensuring that autonomous vehicles can reliably operate in diverse traffic conditions.

3.3. Applications and Improvements of CenterNet in Autonomous Driving

The CenterNet algorithm, through its unique center point detection mechanism, effectively addresses the detection of dense targets and small objects, offering significant advantages for autonomous driving. In practical applications, multiple enhancements to CenterNet have strengthened its performance in various environments.

3.3.1. Center Point Detection Mechanism. CenterNet's center point-based method allows for quick and accurate identification and localization of targets in dense areas. For instance, Zhou et al. showcased this method's superiority across multiple datasets in "CenterNet: Object Detection via Center Point Detection," especially excelling in small object detection [10].

3.3.2. Multi-Modal Data Fusion. Integrating LiDAR and video image data improves CenterNet's target recognition capabilities in complex environments. In "Multi-Object Tracking with CenterNet in Autonomous Driving," researchers combined images and LiDAR information to achieve efficient multi-object tracking, significantly enhancing driving safety [11].

3.3.3. Improved Feature Map Regression. CenterNet introduces a more efficient feature map regression method that optimizes the model architecture, enabling it to handle more complex scenarios. This improvement was further validated in "Center-based Object Detection Network for Autonomous Driving," enhancing detection accuracy and speed [12].

Experimental results demonstrate that the improved CenterNet achieves an Average Precision of 80.1% on the Cityscapes dataset, approximately a 5% increase over its predecessor. In urban traffic environments, CenterNet excels, effectively recognizing pedestrians, bicycles, and other vehicles in congested conditions, thereby enhancing the safety and stability of autonomous driving systems.

4. Experimental Results

Analysis of the above algorithms showcases the application performance of YOLO, RetinaNet, and CenterNet in autonomous driving. Experimental data stem from public datasets (e.g., KITTI and Cityscapes), evaluating metrics such as Average Precision (AP) and time delay. The following table summarizes the performance of these three algorithms and their improvements:

Table 1. Results before and after improvements

As table 1 shown, all three algorithms exhibit excellent performance across different application scenarios. The YOLO series performs exceptionally well in frame rates (FPS), while CenterNet demonstrates the best accuracy (AP), making it particularly suitable for complex dynamic environments. RetinaNet exhibits strong capabilities in small object detection. Furthermore, the improvements to YOLO, RetinaNet, and CenterNet effectively enhance algorithm accuracy. Notably, the optimized versions of these algorithms significantly improve performance in urban traffic environments and small object detection, ensuring the safety and reliability of autonomous driving systems in complex and dynamic settings. These results provide robust theoretical and practical foundations for academics and industries to optimize and develop autonomous driving systems.

5. Challenges and Future Directions

Despite the emergence of many new technologies in the development of object detection algorithms, some unresolved challenges still exist in the field of autonomous driving. First, the complexity of environments, including lighting variations and the diversity of targets ranging from small traffic signals to large vehicles, places immense pressure on detection algorithms. Additionally, the demand for real-time processing continues to rise, making it crucial to optimize the algorithms' speed while ensuring accuracy. Future research can explore the following directions:

Expanding the diversity of training datasets to cover different weather, lighting, road, and traffic conditions will enhance the algorithms' generalization capabilities.

Integrating data from various sensors, including cameras, LiDAR, and radar, to enhance the accuracy and robustness of environmental perception.

Utilizing techniques such as model pruning and quantization to reduce network size and accelerate model inference, making it more suitable for real-time applications.

6. Conclusion

This paper reviews the significance of object detection in the field of autonomous driving, thoroughly examining the basic processes, characteristics, and performance improvements of three algorithms: YOLO, RetinaNet, and CenterNet. Experimental results indicate that all three algorithms demonstrate good performance and potential across various scenarios. While significant progress has been made in object detection within the autonomous driving domain, challenges remain in addressing complex environments and real-time requirements. Future research should focus on enhancing algorithm robustness and generalization capabilities, improving object detection performance in variable environments, and promoting the ongoing development of autonomous driving technology.