Research On Driver Smoking State Recognition Algorithm Based On Optimized Convolution Neural Network

In view of the fact that the smoking behavior of the driver during driving seriously affects driving safety,especially driving vehicles with "two passengers and one danger",the smoking behavior of the driver during driving will have more serious consequences.At present,compared with ordinary non-smoking methods,intelligent monitoring methods have the characteristics of remote monitoring,real-time monitoring,and timely processing.It is an effective method to identify and detect smoking behavior during driving by means of real-time driver image monitoring.Regarding the detection and recognition of smoking behavior,existing researches have focused on detection and recognition based on smoking actions or smoke characteristics.These methods are likely to cause low accuracy of detection and recognition due to some special circumstances,and it is easy to cause false judgment.In view of the above situation,this paper uses the driver behavior images collected by the vehicle platform to improve the recognition and detection of cigarettes in the driver's facial image by improving the Mask R-CNN algorithm,thereby improving the recognition accuracy and reducing the recognition misjudgment rate.This paper first improves the feature extraction part of the algorithm,combines the feature pyramid network and the hollow convolution to construct a new network structure FPN-D,and uses this network structure to classify the experimental comparison of the network structure of the feature extraction part in the original algorithm.The experiment shows that the accuracy rate of the FPN-D feature extraction network in this paper reaches 94.75%,which is 7.25 percentage points higher than that of the original algorithm;the AUC(Area Under the Curve)value of the network is 95.6%,compared with the original algorithm.Of the Internet increased by 4.5 percentage points.Then,according to the characteristics of the data set studied,the anchor frame generation part of the algorithm RPN network is improved to make the algorithm generate anchor frames from the specified pixels.This method not only helps to reduce the number of anchor frames generated during the detection and operation of the algorithm,And it will reduce the time it takes for the algorithm to compare the screening anchor frames,thereby improving the recognition and detection efficiency of the algorithm.Through the improvement of the algorithm,the detection and recognition of the smoking behavior of the driver is realized.Compared with the original algorithm,the improved algorithm can better recognize the object of cigarette in the image,so as tojudge whether the driver has unsafe behavior of smoking while driving.The overall improved experiment of the Mask R-CNN algorithm shows that the accuracy of cigarette recognition in the image is above 94%,and the algorithm can mark the shape of the cigarette recognized by the algorithm,so the improved algorithm can effectively detect and recognize driving Cigarettes in the facial image.