Roadside cameras play a crucial role in road traffic,serving as an indispensable part of integrated vehicleroad-cloud systems due to their extensive visibility and monitoring capabilities.Nevertheless,these cameras fa...Roadside cameras play a crucial role in road traffic,serving as an indispensable part of integrated vehicleroad-cloud systems due to their extensive visibility and monitoring capabilities.Nevertheless,these cameras face challenges in continuously tracking targets across perception domains.To address the issue of tracking vehicles across nonoverlapping perception domains between cameras,we propose a cross-camera vehicle tracking method within a Vehicle–Road–Cloud system that integrates visual and spatiotemporal information.A Gaussian model with microlevel traffic features is trained using vehicle information obtained through online tracking.Finally,the association of vehicle targets is achieved through the Gaussian model combining time and visual feature information.The experimental results indicate that the proposed system demonstrates excellent performance.展开更多
针对鱼眼相机的传统标定过程烦琐并且不适用于日常场景图像的问题,提出了一种新的基于卷积神经网络的方法,可同时标定鱼眼镜头的内参并进行图像畸变校正。该方法通过预测不同畸变参数下像素点的位移量,从而提高鱼眼相机标定和图像畸变...针对鱼眼相机的传统标定过程烦琐并且不适用于日常场景图像的问题,提出了一种新的基于卷积神经网络的方法,可同时标定鱼眼镜头的内参并进行图像畸变校正。该方法通过预测不同畸变参数下像素点的位移量,从而提高鱼眼相机标定和图像畸变校正的精度;为了进一步提高模型精度和泛化性,在编码部分引入坐标注意力模块,增强对图像位置信息的关注度;最后为了增强图像的细节特征,在跨越连接部分设计了跨尺度融合模块。针对数据集稀缺的问题,还生成了一个新的大规模数据集,标有相应的畸变参数和畸变校正后的图像。实验结果表明:与其他鱼眼相机标定方法相比,重投影误差为0.312 pixel,标定的精度较高;与图像畸变处理方法相比,峰值信噪比(peak signal to noise ratio,PSNR)为38.055 dB,结构相似度(structural similarity,SSIM)为0.874,图像畸变校正的质量较好。展开更多
基金the National Natural Science Foundation of China(52172389)Natural Science Foundation of Guangdong Province(2022A1515012080)Tsinghua-Toyota Joint Research Institute Interdisciplinary Program.
文摘Roadside cameras play a crucial role in road traffic,serving as an indispensable part of integrated vehicleroad-cloud systems due to their extensive visibility and monitoring capabilities.Nevertheless,these cameras face challenges in continuously tracking targets across perception domains.To address the issue of tracking vehicles across nonoverlapping perception domains between cameras,we propose a cross-camera vehicle tracking method within a Vehicle–Road–Cloud system that integrates visual and spatiotemporal information.A Gaussian model with microlevel traffic features is trained using vehicle information obtained through online tracking.Finally,the association of vehicle targets is achieved through the Gaussian model combining time and visual feature information.The experimental results indicate that the proposed system demonstrates excellent performance.
文摘针对鱼眼相机的传统标定过程烦琐并且不适用于日常场景图像的问题,提出了一种新的基于卷积神经网络的方法,可同时标定鱼眼镜头的内参并进行图像畸变校正。该方法通过预测不同畸变参数下像素点的位移量,从而提高鱼眼相机标定和图像畸变校正的精度;为了进一步提高模型精度和泛化性,在编码部分引入坐标注意力模块,增强对图像位置信息的关注度;最后为了增强图像的细节特征,在跨越连接部分设计了跨尺度融合模块。针对数据集稀缺的问题,还生成了一个新的大规模数据集,标有相应的畸变参数和畸变校正后的图像。实验结果表明:与其他鱼眼相机标定方法相比,重投影误差为0.312 pixel,标定的精度较高;与图像畸变处理方法相比,峰值信噪比(peak signal to noise ratio,PSNR)为38.055 dB,结构相似度(structural similarity,SSIM)为0.874,图像畸变校正的质量较好。
