Research on Multi-person 3D Human Body Pose Estimation Based on Multi-Branch Collaborative Optimization

With the development of artificial intelligence technology, 3D human pose estimation provides new perspectives in the field of educational examination. However, single-view multi-person 3D pose estimation task in complex scenarios is highly challenging. There is an imbalance in the training process of the multi branch structure method, which limits the overall performance. To solve this problem, we propose a novel single-stage multi-branch estimation method. We use a dynamic adjustment mechanism to consciously reduce the imbalance between tasks, enabling the model to accurately estimate the pose shape parameters while localizing correctly. Compared with state-of-art methods, our method achieves superior performance on standard multi-person benchmarks datasets, such as 3DPW and CMU. In an exam room environment with a large number of candidates and crowded space, it helps to detect and prevent cheating behavior, improving the fairness of the exam.