| In recent years, the number of motor vehicles is increasing in our country, the traffic accidents and casualties is also increasing year by year. Seatbelts is a very important passive protection measures, it can reduce the casualty rate of traffic accidents which cause by vehicles collision or other reasons when vehicles is traveling on the road. Our traffic control department and relevant laws and regulations strictly required the vehicle drivers must to wear seatbelts during driving on the road. But in our country, there were many drivers did not wear seatbelts when they driving, the main reason for this phenomenon is the drivers’ safety awareness is not strong, they have many irregularities to escape the seatbelt reminder system’s detection. Therefore, research on seatbelt detection methods has great importance for raising awareness of drivers to comply with traffic regulations.This paper proposes a seatbelt detection method based on deep learning, this method tried to use the deep learning methods which already have well applications in the area of image recognition to improve the recognition accuracy of seatbelt detection method.Compared to the conventional seatbelt detection methods, the biggest advantage of the deep learning methods is that it can be automatically learning characteristics from sample data, so it can minimize the complexity of manual design features and human intervention.In this method, before using the deep learning method to train and detect, we have to do some preprocessing for transport video images. That is the first use of temporal difference method to obtain the minimum bounding rectangle of the vehicle, then coarse positioning the window of car based on experience thresholds for excluding the interference information cause by the front part of the car. After hat we can use edge detection and integration projection to prescise positioning the window of car, finally we can obtain driver’s window area as a training sample image or detect image. In this paper, we study the convolution neural networks model of deep learning and use this model trained on the sample images in order to get the model for seatbelt detection. Then we can use this model to detect image whether the driver in the image is weared seatbelt. we use Caffe framework to implement thewhole process of training and testing, and the experimental results indicate the proposed method achieves a feasible and valid result for application. |
