Optimizing Large Models for Human Pose Anomaly Detection in Videos

Recent years, with the continuous development of AI technology, video detection image recognition technology is also developing rapidly. In video monitoring and detection systems, human abnormal gesture recognition has received more and more attention. This article is based on the previous research using an improved version of Deepsort, using a new architectural approach, spring using a new training strategy and training dataset, not only using the public dataset, but also combined with the Chinese characteristics of CCTV surveillance video datasets throughout the streets of China for training. Compared with traditional dataset training, our training dataset is more suitable for complex and changeable real scenarios and more in line with the geographical characteristics of Chinese characteristics. In order to make the dataset accurate during the training process, we manually labeled the original dataset, especially for the dataset in China's densely populated environment, which made the dataset more accurate during the training process. Through the adjustment of the optimization parameters of the model and the joint training of multiple data sets, the results of this experiment can prove that after using the improved version of this deep sort, the accuracy and processing speed of human abnormal posture detection have been improved well. Compared with this data from a traditional single data source, the training method of fusing public data and CCTV data can achieve more ideal results.