5results about How to "Reduce matching error" patented technology

This invention provides a method and apparatus for determining a target speaker based on text and lip movement sequence matching. It addresses the problems of low recall, poor robustness, and inaccurate target speaker selection in existing technologies for text-lip movement matching. The method includes: obtaining the corresponding video segment and its duration based on the start and end times of the wake-up word obtained from voice wake-up detection; performing face detection and tracking on the video segment to extract the region of interest (ROI) image sequence of the lips of each candidate speaker in the video segment, as the lip movement sequence to be matched; generating a reference lip movement sequence with the same duration as the video segment based on the wake-up word text; inputting the lip movement sequence to be matched and the reference lip movement sequence into a global-local dual-attention matching network model to calculate the matching score; selecting the lip movement sequence with the highest matching score that is greater than a set threshold, and determining the speaker corresponding to the lip movement sequence to be matched as the target speaker.

The application provides a feature matching method and system based on multi-source heterogeneous data of intelligent terminals, and relates to the technical field of wireless communication. The method comprises the following steps: a target terminal acquires wireless network measurement records and motion state records, and respectively associates time markers; an environment fingerprint formed by at least two types of statistical quantities is generated based on the wireless network measurement records, which is used to indicate the wireless scene characteristics of the observation position; within a continuous time window, a continuity parameter is generated based on the motion state records to represent the motion continuity between records; data objects carrying time markers and associated with the environment fingerprint and the continuity parameter are combined in pairs to form candidate object pairs, and the candidate object pairs come from different modalities or different terminals; based on the environment fingerprint and the continuity parameter, a matching determination is made according to the time markers, and a target matching pair satisfying a predetermined determination rule is output; and the application can improve the accuracy and stability of cross-source data matching in a dynamic indoor scene.

The application discloses a method for determining switching point information between maps, a navigation method, a device, a program product and a medium, and comprises the following steps: when the autonomous mobile device moves from a first sub-map to a transition area, current point cloud data of the environment around the autonomous mobile device is acquired, wherein the transition area represents an overlapping area between the first sub-map and a second sub-map to be entered; based on the current point cloud data, environment feature degradation degree and pose fusion data are calculated, and switching point information of an optimal switching point for entering the second sub-map is dynamically determined. The application can reduce drift when entering one sub-map from another sub-map, and improve the smoothness of map switching of the autonomous mobile device.

The application discloses a wide-area low-altitude non-cooperative target real-time discrimination method based on self-supervised incremental learning, and belongs to the technical field of radar target recognition and air traffic surveillance. Firstly, the radar track cluster of the ground-to-air radar and the cooperative target track cluster of ADS-B are acquired, and the neighborhood context spatio-temporal consistency dynamic time warping method is used for hetero-frequency track similarity measurement and cross-threshold nearest neighbor matching. The matching result is used as self-supervised information to construct a discrimination network based on the Transformer Encoder, the robustness is enhanced through random ordering of the track cluster, and the model is evolved autonomously through incremental learning. Finally, combined with GPU parallel computing, end-to-end fast reasoning with a time complexity of O (N) is realized, and the cooperative / non-cooperative attributes of each radar target are output. The application does not need manual labeling, can accurately match hetero-frequency tracks, significantly improves the real-time performance, accuracy and dynamic adaptability of non-cooperative target discrimination, and can be widely deployed in low-altitude traffic control systems.

This invention discloses a method for building a multimodal fusion-based fault diagnosis model for power transformers, belonging to the field of fault diagnosis model technology. This method deploys multimodal sensors to synchronously acquire transformer vibration and oscillation wave signals. After preprocessing the signals using Empirical Mode Decomposition (EMD), it jointly models the time-domain, frequency-domain, and oscillation wave parameter features. The core of this invention lies in constructing a graph data model based on physical topology and introducing a cross-modal Transformer network for deep feature fusion to capture the complex correlations between different modes. Furthermore, this method utilizes a variational autoencoder (VAE) to achieve unsupervised diagnosis and data augmentation for unknown faults; by constructing a power transformer fault knowledge graph (KG) and using a TransE model for entity embedding, it achieves accurate fault severity assessment and lifespan prediction of power transformers.