A Modified Yolov3 Model (Yolov3sd) for Detection of Drinking and Smoking Actions

Abstract

For many applications of intelligent systems, the ability to detect and recognize objects of interest from images is crucial. Precision and processing efficiency are two key considerations for such an object detection task, particularly for applications which need to operate in real-time. In this paper, a study is carried out to detect smoking and drinking actions from the images, which is very challenging due to various complexities exist in the background of image and must be detected in the real time so that disclaimer can be generated accordingly. To this end, first a dataset is generated by collecting images of smoking and drinking actions from the videos and movies. Then a variant of YOLOv3, a well-known object detector, is proposed named YOLOv3SD, which is the combination of DenseNet and YOLOv3. The DenseNet is combined to retain and reuse the features so that the model can detect actions in complex background with more precision. The experiment results shows that the proposed network has higher detection accuracy compared to the previous state-of-the-art networks such as Faster R-CNN, SSD, YOLOv2 and YOLOv3.