| Summary: | 碩士 === 國立中興大學 === 電機工程學系所 === 107 === Researches and analyses from the US bureau of safety experts reveals that the probability of car accidents for smokers is 1.5 times larger than that for non-smokers. The United Kingdom and Germany experts believe that the 5% of car accidents are related to smoking while driving. According to the US experts, the smoking behavior leads to three driving hazards: reduced vision, distracting, and odors stimulations. People need a high degree of attention and a clear vision while driving, and the smoking behavior may irritate eyes and respiratory tract, and then the vision will be blurred. The acrid smoke also causes coughing and distraction. Specially, the influence of smoke on the vision is very terrible. When a driver is in such circumstances, the risk is very high. Therefore, the development of cigarette detection technology is a very important issue when the driver smoking behavior happens. In this thesis, to overcome the light disturbance and insufficient light conditions in the car by image-based recognition methodologies, the YOLO-based deep learning design is developed to detect the cigarette object when there is the driver smoking condition.
In this thesis, the YOLO V2 based deep learning detector is developed for the cigarette object detection when there is driver smoking behavior. The developed method recognizes the driver smoking behavior in the day and night conditions. To prepare the suitable databases, 28 video sequences of smoking drivers, which have a total of 6000 images are used to label cigarette objects for different driver smoking behaviors, where 600 images are selected for the testing phase and 5400 images are selected for the training phase. In experiments, the applied deep learning based method performs that the precision is up to 97% and the recall is 98%. Besides, by the proposed method, the average accuracy of cigarette detection is up to 96% during the day condition, and that of cigarette detection is up to 85% during the night condition.
|
|---|
