Research on multi-object detection technology for road scenes based on SDG-YOLOv5

Optimization of loss function

The overall loss function of the YOLOv5 detection algorithm combines Confidence Loss, Classification Loss, and Localization Loss, which are weighted and summed together. Specifically, for the bounding box regression loss function, CIoU Loss is utilized in the form below: (1)${\displaystyle L_{CI\text{o}U}=1-I\text{o}U\text{+}\frac{{\rho }^2\left(b,b^{gt}\right)}{c^2}\text{+}\alpha \upsilon }$

where $I\text{o}U=\frac{\left|B\cap B^{gt}\right|}{\left|B\cup B^{gt}\right|}$ represents the intersection ratio between the predicted bounding box B and the actual bounding box B^gt, $\rho \left(b,b^{gt}\right)=d$ is the Euclidean distance separating the center point of B and B^gt, c denotes the diagonal distance of the smallest bounding box that wraps around B and B^gt, α is the weight parameter, which is defined as follows: (2)${\displaystyle \alpha =\frac{\upsilon }{\left(1-I\text{o}U\right)+\upsilon }}$

where υ is the aspect ratio, which is employed to evaluate the conformity of the aspect ratio between B and B^gt. When the center points of B and B^gt coincide, υ can be expressed in the form below: (3)${\displaystyle \upsilon =\frac{4}{{\pi }^2}{\left(arctan\frac{w^{gt}}{h^{gt}}-arctan\frac{w}{h}\right)}^2.}$

CIoU Loss primarily consolidates bounding box regression metrics such as centroid distance, intersection ratio, and aspect ratio. However, it does not explicitly address aligning the orientation between predicted and real bounding boxes. Additionally, when the width and height of bounding boxes B and B^gt converge to $\left\{\left(w=k{w}^{gt},h=k{h}^{gt}\right)\left|k\in R^{+}\right.\right\}$, the width and height of the bounding box regression B cannot simultaneously increase or decrease. This phenomenon leads to delayed convergence of the predicted bounding box during training, thereby hindering the real-time capabilities of the detection method.

SIoU Loss is composed of Angle Loss Λ, Distance Loss, Shape Loss, and IoU Loss. The Angle Loss Λ is defined as follows: (4)${\displaystyle \Lambda =1-2\text{*}{sin}^2\left(arcsin\left(\frac{{\text{c}}_h}{\sigma }\right)-\frac{\pi }{4}\right)=cos\left(2\text{*}\left(arcsin\left(\frac{{\text{c}}_h}{\sigma }\right)-\frac{\pi }{4}\right)\right)}$

where $\sigma =\sqrt{{\left(b_{c_x}^{gt}-b_{c_x}\right)}^2\text{+}{\left(b_{c_y}^{gt}-b_{c_y}\right)}^2}$ and $c_h=max\left(b_{c_y}^{gt},b_{c_y}\right)-min\left(b_{c_y}^{gt},b_{c_y}\right)$; σ represents the distance from the center point of bounding boxes B and B^gt; c_h denotes the disparity in height between the center point of bounding boxes B and B^gt; b_cx, b_cy, $b_{c_x}^{gt}$, $b_{c_y}^{gt}$ are the center coordinates of bounding boxes B and B^gt; $\frac{c_h}{\sigma }$ is the opposite side of the right triangle relative to the hypotenuse $sin\left(\alpha \right)$; In the training process, when $\alpha >\frac{\pi }{4}$, Λ takes β, otherwise, it takes α. The distance Loss can be expressed as: (5)${\displaystyle \Delta =\sum _{\text{t=x},y}\left(1-{\text{e}}^{-\gamma {\rho }_t}\right)=2-{\text{e}}^{-\gamma {\rho }_x}-{\text{e}}^{-\gamma {\rho }_y}}$

where ${\rho }_x={\left(\frac{b_{c_x}^{gt}-b_{c_x}}{c_{w1}}\right)}^2$, ${\rho }_y={\left(\frac{b_{c_y}^{gt}-b_{c_y}}{c_{h1}}\right)}^2$, γ = 2 − Λ,and c_w1, c_h1 are the width and height of the smallest outer rectangle of bounding boxes B and B^gt, respectively. The shape loss is defined in the form below: (6)${\displaystyle \Omega ={\sum _{\text{t=w},h}\left(1-{\text{e}}^{-{\text{w}}_t}\right)}^{\theta }={\left(1-{\text{e}}^{-{\text{w}}_w}\right)}^{\theta }+{\left(1-{\text{e}}^{-{\text{w}}_h}\right)}^{\theta }}$

where ${\text{w}}_{\text{w}}=\frac{\left|\text{w-}{\text{w}}^{gt}\right|}{max\left(\text{w},{\text{w}}^{gt}\right)}$, ${\text{w}}_h=\frac{\left|\text{h-}{\text{h}}^{gt}\right|}{max\left(\text{h,}{\text{h}}^{gt}\right)}$; w, h, w^gt, and h^gt correspond to the width and height of bounding boxes B and B^gt, respectively. θ governs the degree of attention given to shape loss. To prevent an excessive focus on shape loss and minimize the displacement of the prediction bounding box, the value range of θ was set from 2 to 6. SIoU Loss is defined as follows, and its regression principle is shown in Fig. 2. (7)${\displaystyle L_{SI\text{o}U}=1-I\text{o}U\text{+}\frac{\Delta \text{+}\Omega }{2}.}$

Given that the angle loss $cos\left(2\text{*}\left(arcsin\left(\frac{{\text{c}}_h}{\sigma }\right)-\frac{\pi }{4}\right)\right)$ of $arcsin\left(\frac{{\text{c}}_h}{\sigma }\right)-\frac{\pi }{4}=\alpha -\frac{\pi }{4}$ and $\beta =\frac{\pi }{2}-\alpha$ become identical following the cos function, the angle loss is minimized when α = 0 and maximized when it approaches π/4. This facilitates quicker movement of the prediction bounding boxes towards the nearest axis. Consequently, subsequent regression only necessitates X or Y coordinates, leading to a significant improvement in both model detection accuracy and training speed.

The loss function adopted in this article enables faster and more effective alignment of predicted bounding boxes with actual boxes compared to the loss function utilized in the algorithm outlined in Cai et al. (2021). This facilitates expedited training on large datasets, reduces time costs, and enhances the labeling accuracy of predicted bounding boxes on image instances. While the YOLOv6 algorithm employs an anchor-free SIoU Loss, this article utilizes an anchor-based SIoU Loss. Although the anchor-based approach marginally affects detection speed, it notably enhances detection accuracy. In summary, the loss function employed in this study is better suited for intelligent driving application scenarios.

Optimization of the detection head

The original YOLOv5 detection algorithm adopts a coupled head format, as illustrated in Fig. 1, which outputs classification and regression tasks after a convolution layer. However, this approach leads to a conflict between classification, which emphasizes the resemblance of extracted features to specific categories, and regression, which focuses on adjusting predicted bounding box parameters based on differences from actual coordinates. To resolve this problem, Ge et al. (2021) introduced YOLOX with a decoupled head (DH) architecture. This architecture begins with a 1 × 1 convolution to alter the channel number to 256 before branching into two paths for classification and regression, with an additional IoU branch under the regression path. Each path undergoes two 3 × 3 convolution layers and a 1 × 1 convolution layer, resulting in separate heads for classification and regression. Studies Ge et al. (2021) demonstrated that DH enhances training speed and accuracy by resolving task conflict. However, the performed experiments indicate that while DH improves performance, it also increases the parameter count and floating-point operations (Gflops), resulting in higher memory usage and diminished real-time capability of the network. Therefore, this article introduces an enhanced decoupled head, named DHs, which replaces the network’s Head component, as illustrated in Fig. 3.

Figure 3 reveals that the input features H × W × C undergo an initial 1 × 1 convolution to compress the channel number to 256. Subsequently, in both parallel branches, one 3 × 3 convolution is removed to reduce the number of parameters. Finally, a 1 × 1 convolution is employed to extract features for classification and regression tasks. Compared to the coupled head structure of YOLOv5, the classification detection head in the DH improves the prediction of positive samples, while the regression detection head enhances the accuracy of bounding box regression. Ultimately, experimental results demonstrate that the developed DHs, despite reducing a significant number of parameters and floating-point operations, still maintain the advantages of the DH architecture.

In the algorithm outlined in Cai et al. (2021), the number of coupled detection heads was increased from three to five, resulting in a significant enhancement in the precision of predictions for small objects. In contrast, the proposed algorithm replaces three coupled detection heads with decoupled heads, thereby improving the prediction accuracy for small objects while concurrently reducing the number of parameters and the volume of floating-point operations. In comparison to the decoupled heads utilized in the YOLOX algorithm, the decoupled heads in the proposed algorithm maintain detection accuracy while substantially reducing the number of parameters and floating-point operations. Consequently, this accelerates the detection rate and renders it more suitable for practical application scenarios.

Integration of lightweight global attention mechanism

The concept of the attention mechanism originated in the field of visual imagery and gained prominence with an article published by Google DeepMind (Mnih, Heess & Graves, 2014). By integrating the attention mechanism into the image processing workflow of computers, neural networks can emulate the human visual cognitive requirement of rapidly focusing on key information while neglecting less important data. This approach significantly conserves computing resources. Squeeze-and-Excitation Network (SeNet) focuses on the differences in channel information, adaptively adjusting the model based on the attention weights of each channel. However, this can lead to inefficiency issues when suppressing important information (Hu, Shen & Sun, 2018). The Convolutional Block Attention Module (CBAM) serially integrates attention weights in both channel and spatial dimensions on top of the input features, while the Bottleneck Attention Module (BAM) processes channel and spatial dimensions in parallel. However, both overlook the interaction between channel and spatial dimensions, losing cross-dimensional information (Woo et al., 2018; Park et al., 2020). Considering the importance of interactions between cross-dimensional information, the Triplet Attention Module (TAM) aims to enhance network efficiency by leveraging attention weights between channel, spatial width, and height dimensions. Yet, the attention operation is still applied to only two dimensions, not three (Misra et al., 2021). To amplify the interaction between the three dimensions, this article utilizes an attention mechanism called Global Attention Mechanism (GAM) capable of capturing important features in all three dimensions. Similar to CBAM, the overall structure employs a sequential arrangement of channel and spatial attention modules. The distinctions include not only the elimination of pooling operations to preserve as much information as possible but also the incorporation of an MLP to reduce the number of parameters. The specific network structure is depicted in Fig. 4 and mathematically expressed in Eqs. (8) and (9), where F₁ represents the input feature, F₂ denotes the intermediate state, and F₃ is the output feature, while M_c and M_s respectively represent the channel and spatial attention, and ⊗ denotes the element-wise multiplication.

In Fig. 4, the input feature F₁ will initially pass through two branches. The feature of the upper branch will undergo the channel attention module and then multiply with the input feature F₁ of the bottom branch to obtain F₂. Subsequently, F₂ will go through two branches again. The feature of the upper branch will be multiplied with the input feature F₂ of the bottom branch through the spatial attention module to obtain the output feature F₃.

The channel attention module removes the max pooling operation to preserve the integrity of information. It employs a permutation approach to retain information across three dimensions. Subsequently, it utilizes a two-layer multi-layer perceptron (MLP) with an encode-decode structure, having a reduction ratio of γ, to amplify the dependency relationship between channels and spatial dimensions. This process is specifically illustrated in Fig. 5.

In the spatial attention module, to focus on spatial information, two convolutional layers are utilized for the extraction of spatial information. A channel reduction ratio of γ is employed to decrease the number of parameters. This process is schematically illustrated in Fig. 6.

When employing the GAM, a notable drawback arises from the significant increase in the number of parameters due to the two convolutional layers in its spatial attention module. To resolve this issue, the standard convolutions are replaced with group convolutions (GC) and depthwise separable convolutions (DW). The process is schematically illustrated in Fig. 7.

• Group Convolution divides the input feature map and convolution kernels into separate groups, executing convolution operations within these groups to diminish the number of parameters. Suppose the input feature map size is C × H × W and it undergoes N standard K × K convolutions. In that case, the total number of parameters is N × C × K × K. With Group Convolution, the standard convolution is segmented into g groups, with each convolution size being $\frac{C}{g}\times K\times K$, resulting in the total number of parameters being $N\times \frac{C}{g}\times K\times K$. Consequently, the final parameter count is $\frac{1}{g}$ of the original. In this article, the group convolution is divided into four groups, and the modified attention mechanism with group convolution is called GAMGC.

• Depthwise Separable Convolution breaks down a standard convolution into a depthwise convolution and a 1 × 1 pointwise convolution. Unlike standard convolutions that consider semantic information across all channels, each kernel in a depthwise convolution concentrates on a single channel. The size of the depthwise convolution is K × K × 1, resulting in a parameter count of K × K × 1 × N. The pointwise convolution, utilizing a 1 × 1 × C kernel, extracts semantic information between channels, with a parameter count of 1 × 1 × C × N. Thus, the total number of parameters for Depthwise Separable Convolution is K × K × 1 × N + 1 × 1 × C × N, which is even smaller than that of Group Convolution. In this article, the attention mechanism replaced by Depthwise Separable Convolution is termed GAMDW.

In Fig. 7, the two convolutions within the spatial attention module are substituted with DWs and GCs to decrease parameter counts. Experimental analysis demonstrates that these modifications not only reduce the number of parameters but also improve the extraction of essential semantic information. It is worth noting that the performance of GAMGC surpasses that of GAMDW, leading to the selection of GAMGC for incorporation into the YOLOv5 network in this research.

In the algorithm outlined in Zhu et al. (2021), the CBAM attention mechanism is integrated into the neck network of YOLOv5, effectively diminishing the influence of ambiguous information in images on the algorithm’s detection performance. This integration enables the algorithm to concentrate on vital semantic information across two dimensions. Conversely, in this article, the GAMGC is incorporated into both the backbone and neck networks of YOLOv5. This methodology facilitates the capture of critical semantic information across three dimensions, thereby improving target detection accuracy while keeping the parameter count and floating-point operations low.

Experimental dataset & evaluation index & experimental environment

The datasets chosen for this experiment include the Udacity Self Driving Car, BDD100K, and KITTI datasets. The Udacity dataset provides 2D annotations for continuous video images, covering 11 categories: biker, car, pedestrian, traffic light, traffic light-green, traffic light-green left, traffic light-red, traffic light-red left, traffic light-yellow, traffic light-yellow left, and truck. However, to maintain consistency and reduce variability in model results, the category traffic light-yellow left, which has a limited number of labels, was excluded, resulting in 10 categories for the experiment. This dataset comprises a total of 29,800 images with a resolution of 512 × 512, split into training and testing sets in a 9:1 ratio, with 26,579 images for training and 3,221 for testing. The BDD100K dataset consists of 70 k images for training and 10 k for validation. Due to the inadequate number of labels in the train category, this dataset was divided into 12 categories: Person, Rider, Car, Bus, Truck, Bike, Motor, Traffic Light_Green (TL_G), Traffic Light-Red (TL_R), Traffic Light-Yellow (TL_Y), Traffic Light-None (TL_N), and Traffic Sign (TS). The KITTI dataset comprises 7,481 images for training and 7,518 for validation. Since the validation set lacks labels, the 7,481 training images were divided into training and validation sets in an 8:2 ratio. The dataset was categorized into three classes: Car, Pedestrian, and Cyclist. Some images extracted from these databases are depicted in Fig. 8.

The evaluation metrics employed in the experiments include precision (P), recall (R), accuracy (mAP@. 5), number of parameters (Parameters), floating-point operations (GFLOPS), and frame rate (FPS). These metrics are defined as follows: (10)${\displaystyle P=\frac{TP}{TP+FP}}$ (11)${\displaystyle R=\frac{TP}{TP+FN}}$ (12)${\displaystyle mAP@.5=\frac{\sum _{i=1}^{N_{\text{class}}}{\int }_0^1{P}_i{R}_{i}dR}{N_{class}}}$

where TP refers to the number of true positive detections with an IoU greater than a specified threshold, FP refers to the number of false positive detections with an IoU less than or equal to this threshold, and FN indicates the number of undetected true boxes. Furthermore, @.5 denotes an IoU threshold of 0.5, N_class represents the total number of classes, ${\int }_0^1{P}_i{R}_{i}dR$ represents the accuracy of the ith class target, and mAP stands for the mean average precision across all class targets.

The experimental setup and server environment are as follows: the GPU model utilized is an NVIDIA RTX 3090 with 24 GB of VRAM, while the CPU model is an Intel Xeon Platinum 8,350 C. The software stack includes PyTorch version 1.10.0, Python 3.8, and CUDA 11.3, operating on the Ubuntu 20.04 as operating system. The training parameters are configured as follows: the input image size is set to 640 × 640, with a maximum of 300 iterations and a batch size of 32. The optimizer employed is SGD with a momentum of 0.937 and a weight decay coefficient of 0.0005. The initial learning rate is established at 0.01, with dynamic adjustments according to the cosine annealing algorithm, leading to a final learning rate of 0.002.

Analysis of experimental results

In the research conducted, an ablation experiment methodology was employed on the Udacity Self Driving Car dataset, mirroring the approach used by YOLOX. This methodology involved a strategy of parallel insertion of each module, aiming to assess the individual contribution of each module to the overall performance of the model. By utilizing this parallel ablation approach, the study aimed to precisely determine the effectiveness of each module and its impact on fine-tuning the algorithm. This methodology ensures that the effects of each module are evaluated independently, providing a direct means to observe how each component influences the system’s accuracy and efficiency in real-world intelligent driving scenarios. The experimental results are presented in Table 1.

Table 1 indicates that the YOLOv5 algorithm achieves a detection accuracy of 86.1% for road scenes, but with a relatively low recall rate. Conversely, combining SIoU Loss, DHs, and GAMGC algorithms enhances detection accuracy while sustaining the detection rate. Specifically, for road scene detection, the accuracy reached 88.3%, marking a 2.2% increase in mAP@.5.

These enhancements were sequentially integrated into the YOLOv5 network, leading to notable improvements in detection accuracy.

• After integrating the SIoU Loss, there was a notable improvement in various metrics. Specifically, the recall rate, mAP@.5, and the frame rate increased by 1.1%, 0.7%, and 6.9 FPS, respectively. This indicates that the angle loss component within the SIoU Loss plays a crucial role in refining the regression of predicted bounding boxes, particularly benefiting categories with fewer labels. The effectiveness of SIoU Loss may be attributed to its comprehensive design, which accounts for not only the size, position, and aspect ratio of bounding boxes but also their orientation. By considering these geometric properties, SIoU Loss enables a more accurate alignment of predicted boxes with actual ones, especially in scenarios involving closely packed or partially obscured objects. Through better management of angular discrepancies between predicted and actual boxes, SIoU Loss effectively enhances overall precision and detection speed.

• Following the integration of the DH, there was a notable enhancement in various metrics. Specifically, mAP@.5 and the recall rate increased by 1.3% and 12.6%, respectively, indicating a promising accuracy in predicting positive samples and precision in bounding box regression. However, this improvement came at the cost of increased model parameters and computational load, thereby elevating the model’s complexity. To address this challenge, a lightweight version of the decoupled head, termed DHs, was developed. Experimental validation revealed that after incorporating DHs, although there was a slight reduction in mAP@.5 (a decrease of 0.6% compared to +DH, but an improvement of 0.7% compared to the original network), there was a substantial reduction of 24.6% in the model’s parameter count and a decrease of 19.9 GFLOPS in computational load (floating-point operations). Additionally, there was an increase in frame rate by 7.2 FPS. These results demonstrate the capability of DHs to significantly reduce computational costs and enhance processing speed without compromising detection performance significantly, thereby validating the effectiveness and practicality of the optimization strategy employed.

• According to Table 1, the investigation highlights the effectiveness of GAM, GAMGC, and GAMDW methods. The results demonstrate that each method notably enhances the accuracy of the model. GAM utilizes large 7 × 7 convolutions to enhance the model’s ability to capture intricate features. While achieving a 1.8% improvement in mAP@.5 compared to the original network, it also substantially increases the model’s parameters and floating-point operations, thereby impacting the frame rate and detection efficiency. To resolve this challenge, GAMGC employs group convolution, processing input feature maps in groups to reduce the model’s parameters and computational load while retaining sufficient feature learning capabilities. This enhancement enables GAMGC to improve mAP@.5 over the original network while significantly reducing parameters by 56% and floating-point operations by 42.4 GFLOPS, with a substantial frame rate increase of 20.4 FPS. This observation demonstrates GAMGC’s effectiveness in reducing the model’s burden while maintaining detection accuracy. Similarly, GAMDW, by utilizing depthwise separable convolutions, decomposes traditional convolutions into depthwise and pointwise convolutions, further decreasing parameter count and computational complexity. GAMDW achieves a 1.1% mAP@.5 improvement over the original network and outperforms GAM in reducing model parameters by 72.9% and floating-point operations by 55 GFLOPS, with a frame rate increase of 26.3 FPS. The adoption of depthwise separable convolutions not only significantly reduces computational load but also enhances detection accuracy.

While the combination of GAMDW+SIoU Loss+DHs reached an mAP@.5 performance of 87.5%, slightly lower than the combination with GAMGC, this discrepancy may be attributed to GAMDW’s reduction in model parameters and computational complexity through depthwise separable convolutions. Although this design improves computational efficiency, it may not capture complex features as effectively as GAMGC. Consequently, when combined with optimization techniques aimed at improving detection accuracy, GAMDW’s limitations might hinder these techniques from reaching their full potential, resulting in a less pronounced overall performance improvement compared to GAMGC. Given that the performance of GAMDW combined with SIoU Loss and DHs, although superior to the original network, still falls short compared to GAMGC, the choice was made to integrate GAMGC into the YOLOv5 network. This decision allows for more effective utilization of the performance enhancements brought by SIoU Loss and DHs while reducing the computational burden. As a result, high accuracy is achieved while optimizing the model’s computational efficiency and detection speed.

To validate the advantages of the proposed SDG-YOLOv5 over the original YOLOv5 model in terms of detection accuracy for each category on the Udacity Self Driving Car dataset, a category-specific comparative experiment was conducted. This approach aimed to reflect the performance improvements of SDG-YOLOv5 across different categories, allowing for an accurate assessment of the enhanced model’s performance and adaptability in complex scenarios. Through this category-based comparison, the changes in detection accuracy were explored for each category, demonstrating the overall performance enhancement of the model and revealing its strengths and potential areas for improvement in specific categories. The training results of mAP@.5 before and after the improvements are depicted in Fig. 9. In this figure, the average precision (AP) for the categories biker, car, pedestrian, traffic light, traffic light-green, traffic light-green left, traffic light-red, traffic light-red left, traffic light-yellow, and truck are represented by the numbers 1–10, respectively.

Figure 9 reveals notable improvements in detection accuracy across various categories in the enhanced network. Specifically, the ‘biker’, ‘car’, and ‘pedestrian’ categories increased by 2.8%, 0.4%, and 2.7%, respectively. Among the categories related to traffic signals, ‘trafficLight’, ‘trafficLight-Green’, and ‘trafficLight-GreenLeft’ increased by 1%, 2.4%, and 3.2% , respectively, while ‘trafficLight-Red’, ‘trafficLight-RedLeft’, and ‘trafficLight-Yellow’ improved by 2.8%, 2.6%, and 3.9%, respectively. Moreover, it is observed that the detection accuracy of ‘truck’ category is enhanced by 0.6%. These improvements may be attributed to the synergistic effects of several optimization techniques. SIoU Loss, by considering the orientation of bounding boxes, enhanced geometric alignment, leading to improved detection accuracy. The lightweight DHs effectively reduced model complexity while maintaining sufficient feature representation capabilities. Additionally, the attention mechanism (GAMGC) played a crucial role in recognizing small targets like traffic lights by focusing on key features. By incorporating these optimization measures, the SDG-YOLO algorithm achieved performance enhancements across all categories of the Udacity Self Driving Car dataset, significantly improving the accuracy of multi-category target detection.

To evaluate the effectiveness and generalization capability of the proposed SDG-YOLOv5 algorithm, this study conducts comparative experiments on both YOLOv5 and SDG-YOLOv5 algorithms using the BDD100K and KITTI datasets. This experimental approach was chosen because these datasets represent distinct driving scenes and challenges, enabling a comprehensive evaluation of SDG-YOLOv5’s detection accuracy across diverse environments and categories. Through such comparative experiments, the performance of SDG-YOLOv5 can be compared with the original YOLOv5 model in different scenarios. Consequently, the algorithm’s generalization ability can be evaluated, thereby ensuring its reliability and effectiveness in practical applications. Table 2 presents the detection accuracies of SDG-YOLOv5 and YOLOv5 over BDD100K and KITTI datasets.

Table 2 indicates that the mAP@.5 achieved by the YOLOv5 algorithm was 51.9% and 93.6%, respectively. In contrast, the mAP@.5 for the SDG-YOLOv5 algorithm reached 55.3% and 94.6%, representing an improvement in detection accuracy of 3.4% and 1.0% over the original algorithm, respectively. A detailed analysis of the AP values for each category detected by YOLOv5 and SDG-YOLOv5 demonstrates that SDG-YOLOv5 not only achieved significant accuracy enhancements in conventional large target categories such as ‘Car’, ‘Bus’, ‘Pedestrian’, and ‘Cyclist’, but also realized substantial accuracy improvements in small target categories, including different states of traffic lights (‘TL_G’, ‘TL_R’, ‘TL_Y’, ‘TL_N’). These findings effectively validate the efficiency and superiority of the SDG-YOLOv5 algorithm in processing complex traffic scenes, as well as its exceptional generalization capability across datasets.

To demonstrate the superiority of the algorithm proposed in this article, a comparison of its detection accuracy was conducted with current mainstream object detection algorithms on the Udacity Self Driving Car dataset. The experimental results are presented in Table 3. By selecting widely recognized and representative algorithms in this field as benchmarks, the aim is to comprehensively explore the performance advantages of the algorithm in processing complex road scenes. Moreover, this comparison allows for a more objective evaluation of its position and value in the current object detection technology landscape.

Table 3 presents a comparative analysis between SDG-YOLOv5 and existing mainstream algorithms on the Udacity Self Driving Car dataset. In contrast to traditional two-stage algorithms like Faster R-CNN and its enhanced version AD-Faster R-CNN, SDG-YOLOv5 demonstrates an improvement in the mAP@.5 metric by 9.7% and 6%, respectively, along with a boost in frame rate by 19.6 FPS and 24.3 FPS, respectively. This advancement not only demonstrates SDG-YOLOv5’s superiority over traditional two-stage algorithms in effectively integrating spatial and contextual information but also reflects its progress in frame rate, a crucial factor for real-time applications such as intelligent driving. When compared to the single-stage algorithm SSD, SDG-YOLOv5 achieves a notable enhancement of 36.2% in the mAP@.5 metric and an 8.5 FPS increase in frame rate. Additionally, compared to YOLOv3 and YOLOv4-CSPResNeXt, SDG-YOLOv5 respectively exhibits a 7% and 4.9% increase in mAP@.5, demonstrating its effectiveness in improving detection accuracy. In comparison to CF-YOLOX, which incorporates the CBAM attention mechanism, SDG-YOLOv5 has demonstrated a 2.5% enhancement in the mAP@.5 metric, reflecting its effectiveness in improving context-aware features. Additionally, when compared to TPH-YOLOv5, which integrates a self-attention mechanism, SDG-YOLOv5 exhibited a 0.4% increase in the mAP@.5 evaluation metric. While self-attention mechanisms are recognized for improving a model’s feature learning capabilities, they often require longer training periods. The enhancements integrated into SDG-YOLOv5 enable it to notably boost detection accuracy while reducing the training duration, which is particularly vital for applications like intelligent driving requiring rapid iteration and deployment. Against BIGA-YOLO, utilizing lightweight Ghost modules, SDG-YOLOv5 has displayed a 4.2% rise in the mAP@.5 metric. Furthermore, in comparison to the latest YOLOv8, SDG-YOLOv5 showcased a 1% improvement in the mAP@.5 metric, affirming its superior detection accuracy in intelligent driving scenarios. The analysis of the data presented in the table indicates that the proposed algorithm, SDG-YOLOv5, has achieved the highest accuracy while maintaining a frame rate of 30.3 FPS, thus meeting the requirements for real-time detection. Although the detection speed of SDG-YOLOv5 may be slightly slower compared to some mainstream algorithms, prioritizing increased detection accuracy while ensuring real-time performance is deemed more critical in the context of intelligent driving applications.

To evaluate the efficiency of the proposed algorithm, random tests were conducted on the Udacity Self Driving Car dataset using the model loaded with training-exported weights. The test results depicted in Fig. 10 reveal that the original network encounters challenges in accurately detecting certain object categories. Specifically, it demonstrates tendencies to misclassify crucial targets like traffic lights and pedestrians. Moreover, instances of missed detections are observed, where it fails to identify some small and distant objects, thus constraining its utility in traffic scenarios. In contrast, the SDG-YOLOv5 algorithm outlined in this article effectively tackles these issues. It adeptly identifies and precisely localizes targets across diverse road conditions, markedly enhancing the precision of object detection.