Predicting Vehicle Pose in Six Degrees of Freedom from Single Image in Real-World Traffic Environments Using Deep Pretrained Convolutional Networks and Modified Centernet

Autonomous driving is among the top challenges in intelligent transportation systems (ITS) due to its complexities, such as technology reliability, regulatory obstacles, and public acceptance. The scope of autonomous driving in ITS includes enhancing safety, improving traffic efficiency, increasing accessibility, and reducing environmental impact. Different areas of study concentrate on this problem, aiming at big developments in vehicle detection [1], motion prediction [2], and controlling vehicles [3], to help them work better together. Figuring out what is happening in traffic scenes for autonomous vehicles is a major focus, helping them move safely through challenging real-life situations [4]. The critical part about navigating autonomous vehicles is accurately predicting 3D things such as shapes, movement, and translation. To that end, as will be described in the following paragraphs, it is an excellent way of improving one’s ability to predict autonomous vehicle movement when describing traffic scenes with 6D objects positions [5]. A critical issue in autonomous driving is the detection and prediction of vehicle movements based on sensor data [7]. Sensors such as Light Detection and Ranging (LiDAR) [6] devices provide robust and accurate depth measurements of nearby vehicles. However, due to the high cost, interference, and jammer issues in LiDAR, autonomous vehicles use general-purpose commodity cameras. General-purpose commodity cameras also suffer from calibration issues between the two cameras. Therefore, many autonomous vehicles use a hybrid combination of LiDAR, Radio Detection and Ranging (RADAR) [8], and multi-spectral cameras [7].

In this study, we mainly focus on the problem of 3D vehicle pose prediction for surrounding vehicles from a single image in real-world traffic scenarios for an autonomous vehicle. Using a single image as an input to an autonomous vehicle pose prediction model is simpler and less costlier than LiDAR sensors. LiDAR devices contribute to detecting and predicting moving vehicles by emitting laser pulses to accurately measure distance and create 3D maps of the surroundings. However, the high cost of LiDAR technology, susceptibility to interference, and restricted range in unfavorable weather conditions are among the challenges faced by LiDAR devices. This article uses four pretrained deep convolutional networks, ResNet50 [9], ResNext50 [10], Inception-ResNetV2 [11], and DenseNet201 [12], to automatically extract characteristics from 3D car single photos. These characteristics are then used in a modified CenterNet model [13] to predict six-degrees-of-freedom (6DoF) pose, encompassing translation, rotation, and shape.

The current research offers a novel approach for 3D vehicle perception from single input images, involving neighboring vehicles in realistic traffic environments. The proposed solution integrates an end-to-end deep convolutional pretrained network with CenterNet model to detect, segment, and rebuild 3D postures in metric space. The restricted availability of real-world training datasets for the 3D vehicle perception difficulties explains why pretrained deep convolutional networks have been used. Even with limited training datasets, deep convolutional networks outperformed traditional models while decreasing overfitting. Collecting and annotating large-scale training datasets for autonomous vehicles involves significant cost and complexity issues. Leveraging pretrained deep convolutional networks constructed from large ImageNet datasets enables the use of successful transfer learning for 3D vehicle position prediction problems. Transfer learning enhances the model’s prediction abilities, mainly when applied to sparse training datasets.

The performance of the prediction model is crucial for addressing the research problem in autonomous vehicles because even slight misbehavior in real-world traffic environments can lead to significant accidents [14]. Average precision (AP) [15], as with the detection task, is commonly used to evaluate 3D object understanding. However, the similarity is calculated using the 3D bounding box IoU with orientation or the 2D bounding box with viewpoint. The models were then evaluated using mean average precision (mAP). The research results were extensively examined, and the best-performing model for vehicle pose prediction from a single image was selected.

ResNet50, ResNext50, Inception-ResNetV2, and DenseNet201 are the transfer learning models used for automatic feature extraction and are most frequently employed in the literature. Results from these pretrained models are applied to a modified CenterNet model to predict a vehicle’s pose in a dense traffic scenario using a single input image. Understanding 3D objects is critical to the perception system of autonomous vehicles. Traditional bounding box-based addresses used for 3D object comprehension provide significant training challenges, particularly in limiting the effects of repeated detections in the post-processing step. To address these challenges, this study presents a novel technique that leverages center point-based deep convolutional networks to predict the absolute position of each observed vehicle in 3D space.

The main contributions of this work can be summarized as follows:

This research precisely determines nearby vehicles’ spatial orientation (position) using only a single visual image without additional sensory information, through the proposed modified CenterNet model.

Feature extraction serves as a critical precursor in the development of end-to-end deep convolutional networks, which is pivotal in enhancing the model’s ability to discern relevant data from input information. Thus, to provide a more accurate prediction of the vehicle’s actions in complex congested traffic situations, we have decided to utilize four pre-prepared networks for automatic feature extraction. This approach will amplify the Center Point Deep Learning Model’s capabilities and support dependable predictions. Identifying neighboring vehicles and approximating their locations from a single picture is important for smart driving, as it improves safety by avoiding collisions and enables immediate decision making. This method is more cost efficient and reduces hardware complexity compared to using multiple sensors such as LiDAR and RADAR. By merging these advanced techniques, the research aims to push the boundaries of autonomous vehicle perception, ultimately advancing the safety and effectiveness of autonomous driving systems in real-life scenarios.

The remainder of the research paper is organized as follows. The related research in the published literature is covered in Section II. Section III provides a background, architecture, and description of the proposed system. The experimental setup is described in Section IV, along with the datasets utilized for the experiment and the performance indicators used for evaluation. The experimental findings are described in Section V. In Section VI, the model’s performance is discussed with SOTA models. Finally, Section VII concludes with our findings and future directions.

The ability of autonomous cars to correctly anticipate the postures of nearby vehicles based on the detailed 3D features shown in photographs is a critical component of their safe operation [14]. This task involves creating and using sophisticated deep convolutional networks, essential tools in autonomous cars and medical imaging [16], and computer vision [17]. The efficacy of these networks is strongly dependent on the availability of large annotated datasets, as demonstrated by commonly used benchmarks such as ImageNet [18] and MSCOCO [19]. Despite the popularity of such datasets, acquiring enough massive and diverse real-world training data for self-driving vehicles remains a daunting task. For example, while datasets such as KITTI [20] have helped advance our understanding of 3D object perception for autonomous driving, their annotations are sometimes limited, with just a few hundred labeled 3D automobiles. The KITTI database contains sparse annotations, a lack of diversity and irregularities, which may influence the reliability and performance of a driver’s driving. To address this issue, a sizable 3D instance car dataset called “ApolloCar3D” was compiled from images and videos taken in traffic environments in 10 cities in China. Researchers see positive outcomes with supervised or semi-supervised models due to advancements in deep convolutional networks [26]. The SOTA solutions for managing numerous instances by integrating 3D shape models are 3D-RCNN [6] and DeepMANTA [5]. These achievements highlight the critical importance of scientific research in improving the capabilities of autonomous cars. Researchers want to advance the area of autonomous driving by constantly pushing technological frontiers and leveraging the potential of deep learning and computer vision.

Lu et al. [21] introduced a new approach for rebuilding textured 3D models of nearby cars from a single photograph in real-world traffic conditions. This method involves predicting the 6DoF poses of nearby automobiles, which includes their location (x, y, and z) and orientation (roll, pitch, and yaw), based on precise 3D mapping techniques. To this end, the strategy relies on deformable regression and reconstruction algorithms that have proven effective in reproducing objects with complicated shapes. Specifically, the principal component analysis (PCA)-based deformable methods were utilized to simulate the human body. These PCA-based models have been constructed using 78 CAD model files that provide complete representation of most vehicle types seen on the road. But these models are limited in their strength for conventional vehicle types. They have a tendency to find it difficult to properly rebuild larger and more diversified road entities such as trucks, buses, or other rare vehicle types. This restriction is due to the lack of sufficiently diverse models based on PCA in order to represent a large variety of forms and sizes for such larger vehicles. Therefore, Lu et al.’s method is of considerable importance in 3D vehicle reconstruction from one image and underlines the need to further investigate these techniques with a view to incorporating an extensive range of vehicles that can be seen on actual driving environments. Geometric and Scene-aware Network (GSNet) is a tool developed by Ke et al. [22], which estimates 6DoF poses of surrounding vehicles from a single image. GSNet reconstructs 3D vehicle shapes using a divide-and-conquer 3D shape representation strategy with excellent resolution (1352 vertices and 2700 faces). The limitations of GSNet are as follows: it regularizes network training and improves the accuracy of 6D posture estimation. Dense mesh representation forces address geometrical consistency and scene context.

An et al. [23] introduced a deep neural network architecture called “RCBiCenterNet” which is designed to predict the exact position of each object that has been detected in 3D world space. RCBi-CenterNet is a complex recursive composite network, including a Bidirectional Feature Pyramid Network (BiFPN) for successful cross-scale feature fusion and a dual-backbone feature extractor for robust feature extraction. The BiFPN facilitates the rapid integration of data from different scales within RCBiNet, thereby increasing the model’s ability to recognize and analyze objects with various sizes and orientations as part of an image. The dual-backbone feature extractor boosts the model’s potential by leveraging two distinct networks to harvest varied and complimentary information from the input images. The RCBiCenterNet detection head expects that a confidence heatmap will be available to find the location of the detected objects correctly. In addition, the model returns the depth and orientation of each object’s position, providing a thorough understanding of the 3D location and orientation of the objects in the image. However, RCBiCenterNet has a number of serious drawbacks in spite of its strong architecture. The increased complexity of the feature extraction, which mandates that image features are dragged between horizontal and vertical layers, is at the heart of this issue. This additional complexity might lead to a performance reduction, as the substantial processing required can introduce delays and lower the model’s efficiency. In addition, the different layers involved in feature extraction may present potential obstacles to maintain gradient flow during back propagation that could affect network training and convergence.

Transfer learning approaches are critical in improving the prediction capacities of models for self-driving vehicles due to the need and difficulty of gathering large amounts of real-world data [24]. This strategy involves transferring learned weights from past training on large datasets to various artificial intelligence (AI) problems with smaller datasets, which reduces the danger of overfitting [25]. However, different researchers’ use of various deep convolutional networks presents a substantial hurdle in finding the best model [24]. This study takes a comprehensive approach to address this challenging task and close the indicated research gap. It uses pretrained deep convolutional networks to extract features, which are then integrated into a modified CenterNet model. The improved CenterNet design, which includes double convolutional skip connections, is a creative version of the original framework. The double convolution with skip connection provides an alternate path for gradient flow, helping to mitigate the vanishing gradient problem. With a gradual increase in channel complexity, coupled with skip connections, the gradient can be transmitted more easily through the network, which helps stabilize training. Pretrained deep convolutional networks such as ResNet50, ResNext50, Inception-ResNet, and DenseNet201 are replacing the traditional backbone in this enhanced version. The model uses its learned characteristics to enhance the prediction capabilities, when it adds these cutting-edge ideas. This strategic collaboration emphasizes leveraging proven deep learning architectures and transfer learning approaches to improve the performance of prediction models designed for autonomous car applications. By using the abundance of information stored inside pretrained networks, researchers may successfully handle the issues associated with limited real-world data, thereby improving the capabilities and dependability of autonomous driving systems.

The proposed vehicle pose prediction model was applied to the ApolloCar3D dataset. A transfer learning strategy was applied as the backbone network for automatic extracting features by using pretrained weights from ImageNet, and a modified CenterNet model as the head network was used to predict position (x, y, and z) and orientation (roll, pitch, and yaw) for each vehicle present in a single RGB image. In this study, the ApolloCar3D dataset was used. The proposed design of the vehicle pose prediction model is illustrated in Figure 1.

Figure 1:

Schematic design of the proposed modified CenterNet model.

b.

Pretrained networks as backbone

The backbone network of the original CenterNet model is replaced separately with four pretrained networks, namely, ResNet50, ResNext50, Inception-ResNetV2, and DenseNet201, for better time and space efficiency. These networks are already trained on a huge ImageNet dataset, which reduces training time and avoids overfitting problems due to the amount of dataset.

b.i

ResNet50

Residual Network (ResNet) is a Convolutional Neural Network (CNN) design that addressed the “vanishing gradient” problem, allowing networks with thousands of convolutional layers to outperform shallower networks. The shallow CNN with deep layer learns complex features but it leads to vanishing gradient problem. ResNet model introduces residual block with skip connection [27]. Skip connections in ResNet allows different paths to network for gradient to flow. ResNet50 model, as illustrated in Figure 2, consists of 50 layers totally. Input image is passed to 7 × 7 pre-convolutional layer with 64 filters and stride 2, followed by 3 × 3 max pooling layer with stride 2. Then features passed to four subsequent stages. First stage in ResNet50 consists of three residual blocks and each block is made up of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 64, 64, and 256 filters, respectively. Second stage consists of four residual blocks and each block is made up of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 128, 128, and 512 filters, respectively. Third stage consists of six residual blocks and each block is made up of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 256, 256, and 1024 filters, respectively. Final stage consists of three residual blocks and each block is made up of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 512, 512, and 2048 filters, respectively. The network is capable of learning hierarchical features, starting at low levels in the initial stages and moving up to more complex patterns on deeper layers, as a result of an increasing number of filters across each stage. Output of the final stage is passed through average pooling followed by 1000 dimension fully connected network with SoftMax activation function [28].

Figure 2:

ResNet50 architecture.

Each residual block uses a bottleneck design consisting of three convolutional layers. The first layer of each block is a 1 × 1 convolution, which reduces the number of channels, followed by batch normalization and ReLU activation. The second layer is a 3 × 3 convolution, which performs spatial convolution, followed by batch normalization and ReLU activation.

b.ii

ResNext50

ResNext50, as illustrated in Figure 3, is a modified version of ResNet50 with improved performance while reducing network complexity and number of parameters. This is achieved by introducing cardinality (C = 32) in the ResNet50 model [29]. First block consists of 32 cardinalities of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 128, 128, and 256 filters, respectively, and these are repeated three times. Second block consists of 32 cardinalities of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 256, 256, and 512 filters, respectively, and these are repeated four times. Third block consists of 32 cardinalities of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 512, 512, and 1024 filters, respectively, and these are repeated six times. Final block consists of 32 cardinalities of 1 × 1 conv2D, 3 × 3 conv2D, and 1 × 1 conv2D with 1024, 1024, and 2048 filters, respectively, and these are repeated three times. Output of the final residual block is passed through average pooling followed by 1000 dimension fully connected network with SoftMax activation function [30].

Figure 3:

ResNext50 architecture.

Cardinality in ResNext50, with C = 32, splits each block into 32 parallel paths, enhancing the model’s capacity to learn diverse features. It improves performance without significantly increasing computational costs by effectively using parameters. This method ensures a robust gradient flow during training, preventing the disappearance or explosions of these gradients. In addition, the training process for ResNext50 is different from ResNet50 because cardinalities are added to group convolution layers that represent a number of groups. This alteration will increase the parameter space, which may require more computation resources and time for training as well as improve features diversity and representation capability.

b.iii

Inception-ResNetV2

The Inception-ResNet model is a combination of inception architecture and residual connections. The inception module uses a multi-scale convolutional layer applied independently on the input image at every stage using different filter sizes, concatenated, and passed to the next layer [31]. ResNet has deep layers and skip connections to improve performance by avoiding vanishing gradient problems [32]. Inception-ResNetV2 architecture, as illustrated in Figure 4, takes the input image, passes through the stem network, followed by the Inception-ResNet-A block four times, and then passes to the Reduction-A block. Features obtained from the Reduction-A block passed to the Inception-ResNet-B block 10 times, followed by the Reduction-B block. Finally, output passed to the Inception-ResNet-C block five times, followed by average pooling, dropout, and SoftMax activation function. Inception-ResNetV2 employs multi-scale features through inception modules with various kernel sizes, offering a deeper architecture capable of capturing complex patterns in vehicle poses. In scenes of varying occlusions and backgrounds, its efficient design and fine-grained representations make it more accurate.

Figure 4:

Inception-ResNetV2 architecture.

b.iv

DenseNet201

Dense Convolutional Network (DenseNet) is a feed-forward network that connects each layer to other layer. It solves the vanishing-gradient problem, improves feature propagation, promotes feature reuse, and significantly reduces the number of parameters. DenseNet is based on the premise that convolutional networks can be significantly more profound, more accurate, and efficient to train if the connections between layers near the input and those near the output are shorter. DenseNet201, as illustrated in Figure 5, consists of a total of 201 layers. It has four DenseNet blocks, and each layer in the DenseNet block is densely connected with another layer within the respective DenseNet blocks. Features obtained from the last DenseNet block are passed through 7 × 7 global average polling followed by 1000 dimensional fully connected layer and SoftMax activation function [33]. Global average pooling (GAP) aggregates information from all parts of the feature maps, capturing a holistic view of the input image’s features. This helps, regardless of which features are in particular locations, extract the essential information from an entire image. In addition, the GAP algorithm provides a significant decrease in the number of parameters by reducing the size of the features map using an average value per channel. The overall architecture of DenseNet201 demonstrates its use in managing complex data and achieving cutting-edge performance in a wide range of machine-learning applications. DenseNet201 takes a comprehensive approach to deep learning by merging dense connections and unique architectural concepts, significantly improving accuracy, efficiency, and parameter reduction. DenseNet201 is characterized by a dense connection, enabling direct communication between all layers of each block and promoting the reuse of features. To improve model robustness and accuracy, this parameter efficiency facilitates accurate vehicle localization and segmentation through a rich feature hierarchy.

Figure 5:

DenseNet201 architecture.

This section discusses the dataset used, the training process of these deep learning models, and the evaluation criteria used to examine our deep convolutional models’ performance.

a.

Dataset

ApolloCar3D dataset [34], released by Peking University and Baidu Robotics & Autonomous Driving Laboratory, is a fundamental resource for this research. This collection contains almost 60,000 occurrences of 3D car objects, which are carefully formulated from approximately 5277 images using CAD car blueprints. These pictures were taken in four urban centers at peak times, obtaining several velocities and upholding top-notch standards. Particularly, the ApolloScape dataset surpasses the KITTI dataset compared to the quantity of moving objects within it. Moreover, it presents challenging scenarios, including extreme variations in lighting conditions within the same image, such as those induced by shadows cast by overpasses. These challenging scenarios increase computational cost during training of model. Each 4262-image comprising the training dataset contains detailed annotations listing the 3D posture data for every recognized vehicle. These annotations encompass essential parameters, such as yaw, roll, pitch, x, y, and z, wherein (pitch, yaw, and roll) denotes the orientation and rotation, while (x, y, and z) signifies the position and translation [34]. The model type attribute indicates the automotive model associated with a 3D model that may aid in estimating the vehicle’s orientation and size.

As depicted in Figure 6, a center point represents the ground truth vehicle annotations and the bottom rectangle of a 3D bounding box. This depiction provides a clear visualization of the annotated vehicles, facilitating the understanding and evaluating prediction models. The integration of 3D models of vehicles in the dataset provides detailed information about the vehicles position. This integration provides more realistic scenarios for training vehicle pose prediction model. The diverse scenarios captured in urban, suburban, and rural environments in the ApolloCar3D provide generalizability while training model. In addition, the annotations in the database are correct, exact, and compatible with common evaluation metrics such as 3D average precision (A3DP). They provide detailed information of vehicle position in 6DoF, including parameters such as yaw, roll, pitch, x, y, and z.

Figure 6:

Original images with ground truth labels.

b.

Model training

The ApolloCar3D dataset includes 3836 training and 426 testing samples [35]. All the images in the dataset are resized from 3384 × 2710 pixels to 1600 × 700 pixels. Each model was trained for 60 epochs. The learning rate for each model training was 0.001. We then used the AdamW [36] optimizer to optimize mask and regression loss functions. The deep learning algorithms are implemented using PyTorch, and all experiments are conducted on the NVidia K80 GPU. Table 1 and Figure 7 show the respective models’ comparative analysis of training loss.

Figure 7:

Comparative analysis of training loss.

Table 1:

Comparison of training loss

Model	Regression loss	Mask loss	Total loss
Center-Inception-ResNetV2	0.85244	1.6311645	2.484085
Center-ResNet50	0.766475	0.666525	1.432925
Center-DenseNet201	0.450010	2.514673	2.964682
Center-ResNext50	0.466475	0.266525	0.733000

The proposed models underwent rigorous evaluation experiments to assess the impact of different pretrained deep convolutional networks. Results were compared against SOTA models, namely, 3D-RCNN [4] and Direct-Based [34, 36]. Performance evaluation based on A3DP-Abs is presented in Table 2 for the proposed models and Table 3 for the SOTA models. Additionally, performance evaluation based on A3DP-Rel is illustrated in Table 4 for the proposed models and Table 5 for the SOTA models. Evaluation metrics include mAP, with assessments conducted using loose criteria thresholds [2.8:π/6] and strict criteria thresholds [1.3:π/12].

A comparative analysis of Center-ResNet50, Center-ResNext50, Center-Inception-ResNetV2, and Center-DenseNet201, alongside SOTA models, namely, 3D-RCNN and Direct-Based [34], is presented in Figures 8 and 9 based on A3DP-Abs and A3DP-Rel, respectively. The performance metrics of Center-DenseNet201 are as follows: a mAP of 39.92%, a loose criteria-based AP of 52.284%, and a strict criteria-based AP of 44.68% for A3DP-Abs. Correspondingly, for A3DP-Rel, the metrics are 11.82% for mAP, 23.89% for loose criteria-based AP, and 10.85% for strict criteria-based AP. Comparative analysis reveals that the performance based on A3DP-Abs and A3DP-Rel of Center-DenseNet201 surpasses that of Center-ResNet50, Center-ResNext50, and Center-Inception-ResNetV2, as well as the SOTA models, namely, 3D-RCNN and Direct-Based.

Figure 8:

Comparative analysis based on A3DP-Abs. A3DP-Abs, absolute translation thresholds.

Figure 9:

Comparative analysis based on A3DP-Rel. A3DP-Rel, relative translation thresholds.

The advancement of autonomous vehicles owes much to the strides made in AI techniques. The safety and efficiency of autonomous vehicle navigation crucially rely on accurately predicting surrounding vehicles’ 3D attributes, encompassing their translation, rotation, and shape. This becomes particularly critical in congested traffic scenarios where even minor deviations in vehicle behavior can precipitate accidents [14] [36]. This study employs transfer learning within the CenterNet framework to extract features from the ApolloCar3D dataset automatically. By utilizing four pretrained networks—ResNet50, ResNext50, Inception-ResNetV2, and DenseNet201—the aim is to estimate the 6DoF vehicle positions of nearby cars.

Within the existing literature, ResNext50, ResNet50, Inception-ResNetV2, and DenseNet201 emerge as promising and widely used pretrained models for transfer learning. However, to date, comprehensive comparative experiments among these models still need to be included. Given the challenges associated with procuring large labeled training datasets for autonomous vehicles, a transfer learning strategy is adopted for feature extraction to expedite the training process.

Table 1 provides insights into the training loss during the training process of all four models. The performance evaluation of both proposed and SOTA models, based on A3DP-Abs, is depicted in Tables 2 and 3, respectively. Notably, the Center-DenseNet201 model exhibits the highest mAP of 39.92% based on A3DP-Abs, followed closely by Center-ResNet50 with a mAP of 38.85%. Similarly, based on A3DP-Rel, the Center-DenseNet201 model achieves the highest mAP of 11.82% compared to the SOTA models.

Further evaluation based on A3DP-Rel is showcased in Tables 4 and 5 for proposed and SOTA models, respectively. Across both A3DP-Abs and A3DP-Rel performance measures, the Center-DenseNet201 model consistently outshines the other pretrained models alongside SOTA models. These findings underscore the Center-DenseNet201 model’s superior performance in accurately predicting the 6DoF vehicle positions in real-world traffic scenarios.

Using pretrained networks such as ResNet50, ResNext50, Inception-ResNetV2, and DenseNet201 for transfer learning in automotive posture prediction is beneficial. These networks save time and computational resources during training as they have already extracted useful characteristics from large datasets. Since they extract general characteristics applicable to various domains, they can be effectively utilized in new datasets such as ApolloCar3D. Moreover, transfer learning to pretrained networks acts as a form of regularization, preventing overfitting by leveraging knowledge from previous tasks. The performance of pretrained networks is further enhanced by adding two convolutional layers and a skip connection. These skip connections improve gradient flow and facilitate feature reuse, reducing the risk of overfitting on smaller datasets such as ApolloCar3D. They also enable the network to capture more complex patterns and nuances in the data, enhancing its adaptability to the target task. Overall, the efficiency and robustness of transfer learning in automobile posture prediction tasks are generally increased by leveraging pretrained networks with double convolutional layers and skip connections.

A considerable corpus of research on autonomous automobiles emphasizes three core tasks: vehicle recognition, position prediction, and motion control. Accurately anticipating vehicle postures is challenging yet crucial for autonomous cars’ safe navigation in crowded traffic environments. In several AI-based operations, deep dense models such as ResNet50, ResNext50, Inception-ResNetV2, and ConvolutionNet201 have shown promising results. This study utilized transfer learning algorithms to extract features from backbone networks such as ResNet50, ResNext50, Inception-ResNetV2, and DenseNet201. These features were then incorporated into a modified CenterNet architecture to anticipate the postures of surrounding cars based on single photos taken in real-world traffic scenarios, such as headnets. The ApolloScape dataset, notable for its realistic description of diverse driving scenarios and activities, serves as the testing ground for these algorithms. Extensive testing was conducted on the ApolloScape dataset (especially ApolloCar3D), employing four pretrained models: Center-ResNet50, Center-ResNext50, Center-Inception-ResNetV2, and Center-DenseNet201. The Center-DenseNet201 model worked well, reaching a maximum mean average accuracy (mAP) of 39.92% based on A3DP-Abs and a mAP of 11.82% for A3DP-Rel. These findings illustrate the usefulness of the upgraded CenterNet model, coupled with the pretrained DenseNet201 transfer network, in successfully handling vehicle pose prediction challenges.