Cross-Dataset Transfer Driver Expression Recognition via Global Discriminative and Local Structure Knowledge Exploitation in Shared Projection Subspace

摘要

Facial expression is one of the important characteristics of drivers during driving. It is very useful in safe driving detection. Recognizing drivers' expressions by the facial images can be solved with machine learning classification strategies. To obtain a reliable reorganization performance, most of approaches assume that the facial images in the training and testing datasets are independently and identically distributed. However, for real time drivers' facial expression recognition, due to vehicle motion, changes in illumination, noise and head movement, the features displayed for the training dataset may be not valid for the testing dataset. To solve this problem, a novel approach is proposed for cross-dataset transfer driver expression recognition via global discriminative and local structure knowledge exploitation in shared projection subspace (GD-LS-SS). By leaning a shared common subspace, GD-LS-SS utilizes the local geometrical structure of data by exploiting the knowledge of graph topology, meanwhile exploiting the global discriminative information by using the pairwise constrained knowledge between the source and labeled target data. Taking advantage of kernel trick, the kernel version of GD-LS-SS is proposed to learn the kernel projection for handling nonlinear cross-dataset transfer and to further promote the recognition accuracy. Experiments on the KMU-FED dataset show that the satisfactory recognition performance of GD-LS-SS outperforms several traditional non-transfer and related transfer approaches.