Automatic detection of certain unwanted driver behavior
This work is devoted to the automatic detection of unwanted driver behavior such as smoking, using a mobile phone, and eating. The various existing datasets are practically unsuitable for this task. We did not find suitable training data with RGB video sequences shot from the position of the inner mirror. So we investigated the possibility of training the algorithms for this task on an out-of-domain set of people faces images. We also filmed our own test video sequence in a car to test the algorithms. We investigated different existing algorithms working both with one frame and with video sequences and conducted an experimental comparison of them. The availability of temporal information improved quality. Another important aspect is metrics for assessing the quality of the resulting system. We showed that experimental evaluation in this task should be performed on the entire video sequences. We proposed an algorithm for detecting undesirable driver actions and showed its effectiveness.