利用结构化SVM结合CNN的层次化目标检测与人体姿态估计方法

Hierarchical target detection and human body attitude estimation based on structured SVM and CNN

针对现有姿态估计方法不能准确提取特征参数的问题,提出了一种基于结构化支持向量机(SSVM)与卷积神经网络(CNN)的层次化模型。首先,展示了一个基于PS部件模型的SSVM如何实现为一个两层的神经网络,其中第一层是卷积层,另一层是损失增强推理层;通过将模型的结构化形式转换为模型中的一个神经网络,提出方法可以同时学习结构模型和外观模型,同时反向传播误差以学习底层的可学习参数,这些参数可从外观模型特征中提取出来;最后,将SSVM模型转换为神经网络模型,将误差反向传播到较低层,并计算确切的SSVM损失,同时通过基于次梯度的方法来学习原始SSVM。将该模型与当前较为先进的识别模型进行了对比,结果证明提出的层次化模型的识别成功率比对比方法平均高6%,具有更强的识别性能。

Aiming at the problem that the existing attitude estimation method cannot accurately extract the feature parameters,this paper proposed a hierarchical model based on structured support vector machine(SSVM)and convolutional neural network(CNN).Firstly,it showed how a SSVM based on the PS component model could be implemented as a two-layer neural network,where the first layer was the convolutional layer and the other layer was the loss-enhanced inference layer.Then,by transforming the structured form of the model into a neural network in the model,the proposed method could simultaneously learn the structural model and the appearance model,and then backpropagated the error to learn the underlying learnable parameters.These parameters could be extracted from the appearance model features.Finally,it transformed the SSVM model into a neural network model,and propagated the error back to the lower layer,and calculated the exact SSVM loss,while learnt the original SSVM by the sub-gradient-based method.Comparing the model with the current advanced recognition model,the results show that the proposed success rate of the hierarchical model is 6%higher than the comparison method and has stronger recognition performance.