Extending egocentric vision into vehicles: Malaysian dash-cam dataset

Egocentric Vision (EV) has recently become one of the emerging areas in computer vision. It is a process of recording a person’s activity by using a wearable camera installed on the head of that person to gain an exclusive viewpoint. This field has gained more attention recently due to the wide availability of wearable cameras commercially and the recent success of Deep Learning (DL). Such attention has resulted in publishing more EV datasets. In this paper, we argue that the concept of egocentric vision is similar to the recording of road activities by a portable camera installed on the dashboard of a vehicle (commonly known as DashCam) to gain an exclusive viewpoint. We attempt to extend the EV concept into vehicles and show how the two viewpoints are similar. To support this argument, we collect and annotate the first Malaysian DashCam dataset, which consists of 228 video clips about road activities classified into ten categories. We use this dataset to train a DL model on top of VGG16 model -a DL model that is pre-trained on ImageNet-, to perform an initial experiment (activity classification), which reveals some limitations of this work such as small dataset size and less distribution of examples among categories. From this work, we learn that such a dataset may lead to the discovery of valuable statistical knowledge about most frequently occurring behaviors that cause accidents and that we may leverage DL to understand such behaviors to save human life.