MSP-DISK: Naturalistic and Diverse In-Vehicle Database for Joint Pose and Seat Belt Detection

Driver monitoring systems improve the safety of the car by alerting the driver when unsafe behavior is detected. To understand the driver and passengers, it is important to estimate the location of their body keypoints and detect the presence of seat belts. This information can be used to ensure that the driver is facing forward and all passengers are properly wearing their seat belts. Modern computer vision methods perform well in the task of human pose estimation for varied environments. However, in-vehicle scenarios present unique difficulties due to varying lighting conditions, camera placement, and lack of specific in-domain training data. We propose the diverse in-vehicle seat belt and driver keypoint (MSP-DISK) dataset. This corpus consists of images sourced from online video-sharing websites that contain a variety of vehicle interiors, subjects, illumination conditions, camera angles, and camera qualities. Each image contains annotations for the visible upper-body keypoints for the driver and all passengers in the vehicle and a binary mask indicating the location of the seat belt for all subjects. We implement a lightweight computer vision model jointly trained on this dataset to detect seat belts and body keypoint locations. For seat belt detection, we obtain an intersection over union (IoU) equal to 27.8%. For the body keypoint detection, we obtain an Fl-score of 0.773.