非线性化改进的KP-Detector模型在人体姿态识别中的应用

Application of Nonlinear Improved KP-Detector Model in Human Pose Recognition

针对现有体态识别模型检测精度不高、检测周期过长以及模型参数规模过大等问题,提出一种新的改进模型——KP-Detector。该模型将关节点和肢体分开检测和识别,使用改进的PLF匹配方法及Dense连接机制,减少模型复杂度;运用匈牙利算法进行肢体高效匹配,优化使用6层模型结构,同时应用于单人和多人关节点检测。在MPII数据集上测试显示,该模型检测精度优于对比模型,测试速度较其他模型快近4FPS,而模型大小只有18M,具有较大优势。

For existing body recognition models in terms of insufficient detection accuracy,long detection period and too large model parameters,propose a new improved posture recognition model—KP-Detector.The joint points and limbs are detected and identified separately.a new and improved PLF(Point Line Fields)and Dense connection mechanism is used in the model to reduce the complexity of the model and alleviate the disappearance of the gradient of the model.The Hungarian algorithm is used for efficient matching of the limbs,and the 6-layer model structure is optimized.Point positioning phase and two-layer limb detection phase,the model can be applied to single and multi-person joint point detection problems at the same time.On the MII data set,the test accuracy of this model is better than that of the comparison model,the test speed is nearly 4 FPS faster than other models,and the model size is only18 M,which shows that this model has more advantages than other models.