A non-neglectable number of car accidents are caused by driver's loss of ability to drive the car, which may be caused by serious health problems, e.g. heart attack, stroke, drug or alcohol influence, as well as by drowsiness and other problems.
For analysing the human activities, the multi-class deep neural networks are frequently used now, which requires a corresponding training set for each activity. However, the problem that is to be solved here is more complicated, which is mainly due to the fact that the training set can only be obtained for the situations that are normal; for anomaly situations, like serious health problems, the training sets are not available.
The video sequences we proposed can be divided into two main categories. The sequences containing (i) normal driving situations, and (ii) anomalous situations. For training, we have recorded the videos in which the drivers behave normally. As was mentioned before, it is not possible to obtain the records with real serious problems during driving (e.g. heart attack, brain stroke). For testing, therefore, we have captured several records containing anomaly events that were simulated. It is worth mentioning that our data was recorded during driving in a real world and in real traffic, which makes them rare in the area of this topic.
The camera configuration.
An examples of images of the proposed dataset:
