Generation Of Smoking Behavior Dataset Based On Semi-automatic Annotation

In the field of computer vision,human behavior recognition is a branch of its research,in which smoking behavior is a type of human behavior.This behavior is not only harmful to the health of itself and others,but also to human property.Related studies have shown that smoking can cause cancer,heart disease,etc.,and also increase the risk of tuberculosis,certain eye diseases and immune system problems.Smoking is one of the important causes of fires(forest fires,residential building fires,factory fires,etc.).In 2017,there were about 219,000 fires in our country,which caused direct property losses of 2.62 billion yuan,including 15,000 fires caused by smoking.However,in the existing human behavior datasets,there are very few data related to smoking behavior.Currently,they only exist as branches in two large datasets,accounting for less than 2% of the total data.And in the production process of human behavior datasets,mainly through manual annotated,"As much intelligence as there is labor",making the production of datasets expensive and laborious.In order to solve the above two problems,this study collected a lot of smokingrelated data,and labeled them according to the PASCAL VOC labeling guidelines.So as to produce a smoking behavior dataset that can be used in the field of Object Detection.Because of the common dataset production methods existing problem,a semi-automatic labeling scheme was proposed.First of all,two labeling methods are used to manually label smoking data to generate two datasets.Compare and draw close to the large-scale general datasets on each index.Because the amount of manual labeling data grows slowly,I use the proposed semi-automatic labeling scheme to label the smoking behavior data.Then a series of semi-automatic labeling datasets were generated,which greatly reduced the number of manual labeling and had the same properties as the manuallabeling datasets.Finally,the smoking behavior datasets was evaluated with other large-scale general dataset on the existing Object Detection algorithm.Finally produced a smoking behavior dataset DSOT(Dataset for Smok ing Object de Tection)of the same scale and better performance as PASCAL VOC 2007.During the production process,YOLOv3 algorithm was used for semi-automatic annotation,which proved the superiority of semi-automatic annotation.Together with the PASCAL VOC dataset,we test and analysis performance on the two networks of VGG16+Faster R-CNN and Res101+Faster R-CNN,which proved the superiority of the DSOT dataset.