Tsai Bo-Lung, Lin Kwei-Jay
Department of Electrical Engineering and Computer Science, University of California, Irvine, CA 92697, USA.
College of Intelligent Computing, Chang Gung University, Taoyuan 333, Taiwan.
Sensors (Basel). 2024 Mar 21;24(6):2014. doi: 10.3390/s24062014.
In visual-based localization, prior research falls short in addressing challenges for the Internet of Things with limited computational resources. The dominant state-of-the-art models are based on separate feature extractors and descriptors without consideration of the constraints of small hardware, the issue of inconsistent image scale, or the presence of multi-objects. We introduce "OOPose", a real-time object-oriented pose estimation framework that leverages dense features from off-the-shelf object detection neural networks. It balances between pixel-matching accuracy and processing speed, enhancing overall performance. When input images share a comparable set of features, their matching accuracy is substantially heightened, while the reduction in image size facilitates faster processing but may compromise accuracy. OOPose resizes both the original library and cropped query object images to a width of 416 pixels. This adjustment results in a 2.4-fold improvement in pose accuracy and an 8.6-fold increase in processing speed. Moreover, OOPose eliminates the need for traditional sparse point extraction and description processes by capitalizing on dense network backbone features and selecting the detected query objects and sources of object library images, ensuring not only 1.3 times more accurate results but also three times greater stability compared to real-time sparse ORB matching algorithms. Beyond enhancements, we demonstrated the feasibility of OOPose in an autonomous mobile robot, enabling self-localization with a single camera at 10 FPS on a single CPU. It proves the cost-effectiveness and real-world applicability of OOPose for small embedded devices, setting the stage for potential markets and providing end-users with distinct advantages.
在基于视觉的定位中，先前的研究在应对计算资源有限的物联网挑战方面存在不足。主流的先进模型基于单独的特征提取器和描述符，没有考虑小型硬件的限制、图像尺度不一致的问题或多物体的存在。我们引入了“OOPose”，这是一个实时面向对象的姿态估计框架，它利用现成的目标检测神经网络的密集特征。它在像素匹配精度和处理速度之间取得平衡，提高了整体性能。当输入图像共享一组可比的特征时，它们的匹配精度会大幅提高，而图像尺寸的减小有助于更快地处理，但可能会影响精度。OOPose将原始库图像和裁剪后的查询目标图像都调整为宽度416像素。这种调整使姿态精度提高了2.4倍，处理速度提高了8.6倍。此外，OOPose通过利用密集的网络主干特征并选择检测到的查询目标和目标库图像的来源，消除了对传统稀疏点提取和描述过程的需求，与实时稀疏ORB匹配算法相比，不仅确保了结果的准确性提高1.3倍，而且稳定性提高了三倍。除了这些改进，我们还展示了OOPose在自主移动机器人中的可行性，使其能够在单个CPU上以10帧每秒的速度通过单个摄像头进行自我定位。这证明了OOPose对于小型嵌入式设备的成本效益和实际适用性，为潜在市场奠定了基础，并为终端用户提供了明显的优势。
