The real-time monitoring of a driver's gaze region is essential for human-machine shared driving vehicles to understand and predict the driver's intentions. Because of the limited computational resources and storage capacity of in-vehicle platforms, existing gaze region estimation algorithms often hardly balance accuracy and real-time performance and ignore temporal information.

Therefore, this paper proposes a lightweight spatial feature encoding network (LSFENet) for driver gaze region estimation. First, the image sequence of the driver's upper body is captured by an RGB camera. Image preprocessing steps, including face alignment and glasses removal, are performed to obtain left- and right-eye images and facial keypoint coordinates to handle challenges such as cluttered backgrounds and facial occlusions in the captured images. Face alignment is conducted using the multi-task cascaded convolutional network algorithm, and the glasses are removed using the cycle-consistent adversarial network algorithm. Second, we build the LSFENet feature extraction network based on the GCSbottleneck module to improve the MobileNetV2 architecture, since the inverted residual structure in the MobileNetV2 network requires a significant amount of memory and floating-point operations and ignores the redundancy and the correlation among the feature maps. We embed a ghost module to improve memory consumption and integrate the channel and spatial attention modules to extract the cross-channel and spatial information from the feature map. Next, the Kronecker product is used to fuse eye features with facial keypoint features to reduce the impact of the information complexity imbalance. Then, the fused features from the images at continuous frames are input into a recurrent neural network to estimate the gaze zone of the image sequence. Finally, the proposed network is evaluated using the public driver gaze in the wild (DGW) dataset and a self-collected dataset. The evaluation metrics include the number of parameters, the floating-point operations per second (FLOPs), the frames per second (FPS), and the F₁ score.

The experimental results showed the following: (1) The gaze region estimation accuracy of the proposed algorithm was 97.08%, which was approximately 7% higher than that of the original MobileNetV2. Additionally, both the number of parameters and FLOPs were reduced by 22.5%, and the FPS was improved by 36.43%. The proposed network had a frame rate of approximately 103 FPS and satisfied the computational efficiency and accuracy requirements under in-vehicle environments. (2) The estimation accuracies of the gaze regions 1, 2, 3, 4, and 9 were over 85% for the proposed algorithm. The macro-average and micro-average precisions of the DGW dataset reached 74.32% and 76.01%, respectively. (3) The proposed algorithm provided high classification accuracy for fine-grained eye images with small intra-class differences. (4) The visualization results of the class activation mapping demonstrated that the proposed algorithm had strong adaptability to various lighting conditions and glass occlusion situations.

The research results are of great significance for the recognition of a driver's visual distraction states.