基于内卷神经网络的轻量化步态识别方法

Lightweight gait recognition method based on involution neural network

现有步态识别方法存在计算量大、识别速率较慢和易受视角变化影响等弊端,会造成模型难以部署、步态识别准确率降低等问题。针对以上问题本文提出一种基于内卷神经网络的高准确率步态识别方法。首先,基于残差网络架构和内卷神经网络算子提出了内卷神经网络模型,该模型利用内卷层实现步态特征提取以达到减少模型训练参数的目的;然后,在内卷神经网络模型基础上,建立一个由三元组损失函数和传统损失函数Softmax loss组成的联合损失函数,该函数使所提出的模型具有更好的识别性能及更高的跨视角条件的识别准确率;最后,基于CASIA-B步态数据集进行实验验证。实验结果表明,本文所提方法的网络模型参数量仅有5.04 MB,与改进前的残差网络相比参数量减少了53.46%;此外,本文网络在相同视角以及跨视角条件下相比主流算法具有更好的识别准确率,解决了视角变化情况下步态识别准确率降低的问题。

The existing gait recognition methods have disadvantages such as heavy computation,slow recognition rate and easily being affected by the angle of view changes,which makes it difficult to deploy the model and reduces the accuracy of gait recognition.This paper proposes a high accuracy gait recognition method based on involution neural network to solve the above problem.Firstly,an involution neural network model based on residual network architecture and involution neural network operator is proposed,in which the model uses the entrainment layer to extract gait features to reduce model training parameters.Then,based on the involution neural network model,a joint loss function consisting of Triplet loss and traditional loss function(Softmax loss)is established.The function makes the proposed model have better recognition performance and higher recognition accuracy across view conditions.Finally,experimental verification is carried out based on CASIA-B gait dataset.The experimental results show that the number of parameters of the proposed method is only 5.04 MB,which is reduced by 53.46%compared with the residual network before modification.In addition,the proposed network has better recognition accuracy than the mainstream algorithm under the same angle of view and cross-view conditions,solving the problem of reduced gait recognition accuracy under the angle of view changes.