Abnormal Search Action Recognition Dataset in Home

The current security industry faces challenges in home security monitoring, especially for detecting abnormal behaviors of visitors, nannies, etc. Existing monitoring systems mostly rely on static image analysis, which cannot effectively identify dynamic behaviors, leading to potential security risks not being detected in time. This dataset aims to improve the accuracy of abnormal action recognition through high-quality video data, helping users detect potential security issues in a timely manner. Data collection utilizes high-resolution cameras and is conducted in various home environments to ensure coverage of different lighting and scene changes. To ensure data quality, multiple rounds of labeling and expert review have been conducted to ensure labeling consistency and accuracy. Data storage uses MP4 format, with a clear structure for easy subsequent processing and analysis. The core advantage of this dataset is its high labeling accuracy, with labeling consistency above 95% and high data integrity. We have applied new data augmentation techniques, increasing model robustness by 20%. Models trained with this dataset can improve the detection accuracy of abnormal behaviors to 90%, effectively solving practical problems in home security.