Lighter, Smarter, Faster: New AI Model Boosts EV Charging Port Detection
In the rapidly evolving landscape of electric mobility, one of the most critical yet overlooked challenges lies not in battery technology or charging speed—but in the final, precise step of connecting the charger to the vehicle. As demand for fully autonomous charging solutions grows, the ability of machines to accurately and reliably identify a car’s charging port under real-world conditions has emerged as a pivotal technological hurdle. Now, a team of researchers from Hebei University of Science and Technology has unveiled a breakthrough in artificial intelligence that could accelerate the arrival of truly hands-free EV charging.
Led by Dr. Zhao Xiaodong, along with colleagues Liu Ruiqing, Wang Xiang, and Wen Shitao, the research introduces a significantly enhanced version of the YOLOv5 object detection algorithm, specifically tailored for identifying electric vehicle charging ports in complex and variable environments. Their work, published in the Journal of Chongqing University of Technology (Natural Science), presents a model that not only outperforms existing methods in accuracy but also achieves a lighter footprint—making it ideal for deployment in embedded systems such as robotic charging arms or smart parking infrastructure.
The challenge of automated charging port recognition is deceptively complex. Unlike high-contrast, well-lit laboratory conditions, real-world scenarios involve a host of variables: glare from direct sunlight, shadows in underground parking garages, obscured or partially blocked ports, and vehicles parked at awkward angles. In such conditions, even state-of-the-art vision systems can falter, leading to failed connections, system delays, or potential damage to the port or charging equipment. Traditional object detection models, while powerful, often struggle with small-scale targets and are computationally heavy—two factors that severely limit their practicality in autonomous charging applications.
Recognizing these limitations, the research team set out to refine the YOLOv5 architecture, a popular choice for real-time detection due to its balance of speed and precision. However, they identified several areas where the standard model could be optimized for the unique demands of charging port detection. Their approach was not to reinvent the wheel, but to thoughtfully enhance key components of the network, resulting in a more robust, efficient, and accurate system.
At the heart of their innovation is a multi-pronged strategy that addresses the core weaknesses of the original model. First, they replaced the standard Feature Pyramid Network (FPN) with a Bidirectional Feature Pyramid Network (BiFPN). This change may sound technical, but its impact is profound. In deep learning, different layers of a neural network capture different types of information—shallow layers pick up fine details like edges and textures, while deeper layers understand broader context and semantics. The original FPN passes information in one direction, which can lead to the loss of critical detail, especially for small objects like a charging port viewed from a distance. BiFPN, by contrast, enables two-way communication between layers, allowing both fine-grained details and high-level context to be preserved and fused more effectively. This bidirectional flow ensures that even when a charging port appears as a tiny, dimly lit feature in a wide-angle shot, the model can still detect it with confidence.
To further enhance the model’s ability to focus on what matters, the team integrated the SENet (Squeeze-and-Excitation Network) attention mechanism into the backbone. Attention mechanisms, inspired by human visual perception, allow a model to dynamically prioritize the most relevant parts of an image. In the context of charging port detection, this means the network can learn to emphasize the distinctive shape, texture, and color of the port while suppressing irrelevant background elements—such as reflections on a car’s body, nearby signage, or adjacent vehicles. By adaptively weighting the importance of different feature channels, SENet helps the model become more discriminative, reducing false positives and improving overall detection reliability.
Another major contribution of the study is the adoption of GhostNet as the model’s backbone, replacing the original CSPDarknet. This shift is crucial for real-world deployment. GhostNet is designed with efficiency in mind, using a technique called “Ghost modules” to generate more features from fewer computations. Instead of applying expensive standard convolutions across the entire input, GhostNet first uses a small number of primary convolutions and then applies lightweight transformations to create additional, “ghost” features. This approach dramatically reduces the number of parameters and computational operations required, resulting in a much leaner model. For applications like autonomous charging, where processing power and energy consumption are constrained—especially in mobile or embedded systems—this efficiency is not just beneficial, it’s essential.
The final piece of the puzzle lies in the model’s loss function, which guides the learning process by measuring how far the model’s predictions are from the ground truth. The original YOLOv5 uses CIoU (Complete Intersection over Union), a sophisticated metric that considers overlap, distance, and aspect ratio. While effective, CIoU can sometimes struggle with precise boundary regression, particularly when the aspect ratio of the predicted and actual bounding boxes differs significantly. To address this, the researchers introduced EIoU (Efficient IoU) loss, which decomposes the aspect ratio component into separate width and height terms. This allows for more direct and accurate optimization of the box dimensions, leading to tighter, more precise bounding boxes around the detected charging ports. This improvement is particularly valuable in robotic applications, where millimeter-level accuracy can determine whether a charging nozzle docks successfully or misses its target.
The results of this comprehensive redesign are compelling. When tested on a custom dataset of 3,200 images—carefully collected from diverse urban charging stations and augmented with techniques like brightness and contrast adjustment to simulate challenging conditions—the improved model achieved a mean Average Precision (mAP) of 94.75%. This represents a significant leap from the original YOLOv5’s 89.7%, demonstrating a clear gain in detection accuracy. More impressively, the model’s size was reduced from 13.7 MB to just 6.76 MB—a reduction of 6.94 MB—without sacrificing speed. The model still operates at 122 frames per second (FPS), well within the range needed for real-time processing in dynamic environments.
To put these numbers into perspective, the team compared their model against several leading object detection frameworks, including Faster R-CNN, YOLOv3, and SSD. While Faster R-CNN achieved a respectable 60.4% mAP, its processing speed of 46.7 FPS and large model size (84.7 MB) make it impractical for real-time embedded use. YOLOv3 and SSD performed better, with mAPs of 84.3% and 80.6% respectively, but both fell short of the new model’s accuracy. Crucially, the proposed method outperformed the original YOLOv5 in every key metric: higher accuracy, smaller size, and competitive speed. This combination of attributes makes it uniquely suited for the demanding requirements of autonomous EV charging.
The implications of this research extend beyond the laboratory. As cities and private operators invest in smart charging infrastructure, the ability to offer truly automated, touchless charging will be a key differentiator. Imagine a scenario where a driver parks in a designated spot, activates the charging system via an app, and a robotic arm extends from the charging station, locates the port with pinpoint accuracy, and connects the cable—all without human intervention. Such systems are already being piloted in select locations, but their widespread adoption has been hindered by reliability issues. This new AI model addresses those concerns head-on, offering a level of robustness that could make autonomous charging a mainstream reality.
Moreover, the model’s lightweight nature opens the door to deployment in a variety of form factors. It could be integrated into compact, low-power edge devices mounted on charging robots, or run on the onboard computers of autonomous valet parking systems. Its ability to handle diverse lighting conditions and small target sizes means it can function effectively in both sun-drenched outdoor lots and dimly lit underground garages—environments where many current systems struggle.
The research team also conducted a series of ablation studies to validate the contribution of each modification. These experiments systematically added or removed components—BiFPN, SENet, GhostNet, and EIoU loss—to isolate their individual effects. The results confirmed that each element plays a vital role: BiFPN alone boosted mAP by 1.59%, SENet helped reduce model size while maintaining accuracy, GhostNet delivered the largest gains in both efficiency and performance, and EIoU provided the final, crucial refinement in localization precision. This rigorous validation underscores the scientific rigor behind the work and provides a clear roadmap for future optimizations.
Looking ahead, the authors suggest several avenues for further development. One is to expand the dataset to include a wider variety of vehicle models, charging port designs, and environmental conditions. Another is to explore multi-camera setups, using stereo vision or multiple angles to improve depth perception and 3D localization of the port. Ultimately, the goal is to transition from 2D detection to full 6D pose estimation—knowing not just where the port is, but also its orientation in space, which is essential for robotic docking.
The work of Zhao Xiaodong, Liu Ruiqing, Wang Xiang, and Wen Shitao represents a significant step forward in the quest for seamless, automated EV charging. By thoughtfully combining advances in multi-scale feature fusion, attention mechanisms, network efficiency, and loss function design, they have created a model that is not only more accurate but also more practical for real-world deployment. As the automotive industry continues its shift toward electrification and autonomy, innovations like this will play a crucial role in shaping the user experience—turning what was once a manual, sometimes frustrating task into a smooth, invisible part of daily life.
In a world where convenience and reliability are paramount, the ability to simply park and walk away, knowing your car will charge itself, is no longer a distant dream. Thanks to this new AI-powered detection method, that future is drawing closer.
Zhao Xiaodong, Liu Ruiqing, Wang Xiang, Wen Shitao, Hebei University of Science and Technology, Journal of Chongqing University of Technology (Natural Science), doi: 10.3969/j.issn.1674-8425(z).2024.07.015