Video-based peripheral oxygen saturation (SpO2) estimation, utilizing solely RGB cameras, offers a non-contact approach to measuring blood oxygen levels. Previous studies set a stable and unchanging environment as the premise for non-contact blood oxygen estimation. Additionally, they utilized a small amount of labeled data for system training and learning. However, it is challenging to train optimal model parameters with a small dataset. The accuracy of blood oxygen detection is easily affected by ambient light and subject movement. To address these issues, this paper proposes a contrastive learning spatiotemporal attention network (CL-SPO2Net), an innovative semi-supervised network for video-based SpO2 estimation. Spatiotemporal similarities in remote photoplethysmography (rPPG) signals were found in video segments containing facial or hand regions. Subsequently, integrating deep neural networks with machine learning expertise enabled the estimation of SpO2. The method had good feasibility in the case of small-scale labeled datasets, with the mean absolute error between the camera and the reference pulse oximeter of 0.85% in the stable environment, 1.13% with lighting fluctuations, and 1.20% in the facial rotation situation.
Simultaneous localization and mapping (SLAM) is one of the core technologies for intelligent mobile robots. However, when robots perform VSLAM in dynamic scenes, dynamic objects can reduce the accuracy of mapping and localization. If deep learning-based semantic information is introduced into the SLAM system to eliminate the influence of dynamic objects, it will require high computing costs. To address this issue, this paper proposes a method called YF-SLAM, which is based on a lightweight object detection network called YOLO-Fastest and tightly coupled with depth geometry to remove dynamic feature points. This method can quickly identify the dynamic target area in a dynamic scene and then use depth geometry constraints to filter out dynamic feature points, thereby optimizing the VSLAM positioning performance while ensuring real-time and efficient operation of the system. This paper evaluates the proposed method on the publicly available TUM dataset and a self-made indoor dataset. Compared with ORB-SLAM2, the root-mean-square error of the Absolute Trajectory Error (ATE) can be reduced by 98.27%. The system successfully locates and constructs an accurate environmental map in a real indoor dynamic environment using a mobile robot. It is a VSLAM system that can run in real-time on low-power embedded platforms.