Kai Zhang

Camera relocalization plays a vital role in the realms of machine perception, robotics, and augmented reality. Direct learning methods based on structures can have a learning-based approach that can learn scene coordinates and use them for camera position estimation. However, the two-stage learning of scene coordinate regression and camera position estimation can result in some of the scene coordinate regression knowledge being lost throughout the learning process of the final pose estimation system, thereby reducing the accuracy of the pose estimation. This paper introduces an innovative end-to-end learning framework tailored for visual camera relocalization by employing both RGB and RGB-D images. Distinguished by its integration of scene coordinate regression with pose estimation into a concurrent inner and outer loop during a singular training phase, this framework notably enhances pose estimation accuracy. Engineered for flexibility, it accommodates training with or without depth cues and necessitates merely a single RGB image during testing. Empirical evaluation substantiates the proposed method’s state-of-the-art precision, attaining an average pose accuracy of 0.019 m and 0.74° on the indoor 7Scenes dataset, together with 0.162 m and 0.30° on the outdoor Cambridge Landmarks dataset.

With the popularity and portability of smartphones, it is feasible to continuously measure and monitor people’s daily activities with a variety of sensors built into smartphones, such as health diagnosis and motion detection. Due to the flexibility of smartphones in practice, the acquisition of data from sensors with position and orientation effects leads to a decrease in the accuracy of activity recognition. In this paper, we propose a robust human activity recognition system based on extended Kalman filtering (EKF) with adaptive local and global principal component analysis (ALGPCA), temporal channel interaction and attention neural network (TCIANet). The EKF-ALGPCA sensor data preprocessing method is proposed for eliminating the effects of orientation and position changes. And an efficient human activity recognition neural network model TCIANet is proposed, in which the temporal features of collected data can also be extracted during feature extraction using time channel interaction module in the convolutional module, and the temporal attention module is used after the convolutional feature extraction to deal with the long-time dependence of temporal data. The experimental data collected with two kinds of smartphones, Xiaomi and Huawei, in different orientations and positions, are used to prove the effectiveness of the modification method in continuous behavior recognition under the condition of changing orientation and position of smartphones.

The use of remote sensing imagery has become indispensable for long-term crop mapping. Synthetic aperture radar can provide high-resolution polarization information on crop time-series canopy changes, making large-scale, long-time-series crop mapping possible. This study presents a method for precise crop mapping using Sentinel-1 data time series from 2015 to 2020. The proposed method, known as the temporal-spatial fusion U-Net (TSFUNet) model, makes use of the temporal characteristics of the time series to improve crop mapping precision. The model consists of a temporal channel fusion module for combining the temporal characteristics of multiple channels. In addition, an improved U-Net with a cascade feature fusion module is introduced to fuse spatial contextual information at different scales for the generation of final crop segmentation. Experimental results demonstrate that the TSFUNet model for temporal Sentinel-1 sequence-based crop mapping outperforms other state-of-the-art methods. For crop mapping at different times, a multi-stage learning (MSL) training strategy is proposed to improve the model’s generalization ability. In addition, the chaotic-Jaya algorithm is employed to optimize the selection of learning rate hyperparameters in MSL and further enhance the model’s high-accuracy crop mapping performance.

Location information is the core data in IoT applications, which is the essential foundation for scene interpretation and interconnection of everything, and thus high-precision positioning is becoming an immediate need. However, the non-line-of-sight (NLOS) effect of indoor complex environment on UWB signal occlusion has been a major factor limiting the improvement in ultra-wideband (UWB) positioning accuracy, and the optimization of NLOS error has not yet been studied in a targeted manner. To this end, this paper deeply analyzes indoor scenes, divides NLOS into two forms of spatial occlusion and human occlusion, and proposes a particle filtering algorithm based on LOS/NLOS mapping and NLOS error optimization. This algorithm is targeted to optimize the influence of two different forms of NLOS, using spatial a priori information to accurately judge the LOS/NLOS situation of the anchor, optimizing the NLOS anchor ranging using IMU to project the virtual position, judging whether the LOS anchor is affected by human occlusion, and correcting the affected LOS anchor using the established human occlusion error model. Through experimental verification, the algorithm can effectively suppress two different NLOS errors of spatial structure and human occlusion and can achieve continuous and reliable high-precision positioning and tracking in complex indoor environments.