Abstract
Drowsy driver alert systems have been developed to reduce and prevent car accidents. Existing vision-based systems are usually restricted to using visual cues, and they usually depend on tedious parameter tuning or cannot work under general conditions. One additional crucial issue is the lack of public datasets that can be used to evaluate the performance of different methods. In this thesis, we develop two novel systems, i.e. a Component-wise Discretized Deep Belief Network (CDDBN) system and a novel Hierarchical Temporal Deep Belief Network (HTDBN) system, for drowsy driver detection. In CDDBN, after detecting drowsiness-related symptoms using traditional DHMMs and SVM, detailed facial feature are computed to construct several discretized components. The input visible units for DBN are formed by the average of the discretized vectors over a time duration and the softmax layer at the last hidden layer of DBN is to predict the level of drowsiness. In HTDBN, our scheme first extracts high-level facial and head feature representations and then uses them to recognize drowsiness-related symptoms. Two discrete-hidden Markov models that utilize a hash-based scheme are constructed on top of the DBNs. They are used to model and capture the interactive relations between eyes, mouth and head motions. Finally, the summed difference of DHMM likelihoods is used to determine the drowsiness level. To evaluate the performance of the drowsy driver detection systems, we also collect a large comprehensive video dataset containing driver videos of various ethnicities, genders, lighting conditions and driving scenarios. Experimental results demonstrate the feasibility of the proposed CDDBN and HTDBN framework for detecting drowsiness based on different visual cues.