Research On Object Detection Algorithm In Smoking Scene Based On Improved YOLOv5

Smoking has been banned in special places such as petrol stations,production workshops and forestry protection areas because it creates a fire hazard and causes incalculable damage to society and individuals.With the continuous development of technology,automated smoking detection systems can replace manual monitoring and control of smoking behavior.Commonly used automated smoking detection systems can be divided into smoke-based detection,smoking gesture recognition-based detection and cigarette-based detection.Smoke based smoking detection systems are less accurate in outdoor or well-ventilated indoor environments.Smoking gesture-based detection is less capable of detecting different individuals due to factors such as individual differences and postural variations.Smoke based detection can be divided into two types of smoking detection systems based on traditional methods and deep learning,where smoke detection based on traditional methods has the problem of poor accuracy and difficulty in coping with complex scenes.Therefore,in order to solve the above problems,this paper adopts a deep learning-based cigarette detection method and improves the YOLOv5 algorithm to build an efficient and accurate smoking detection system,specifically around the following aspects of research.(1)For the selected benchmark model,the YOLOv5 model was scaled,and the scaled five models were trained and tested on the COCO dataset.The YOLOv5 s were finally selected as the benchmark model for improvement based on detection accuracy,number of model parameters and running speed.(2)In the backbone network part of the model,the Contextual Transformer(COT)module is used instead of the original C3 module to make full use of the contextual information between the input keys to guide the learning of the dynamic attention matrix,thus enhancing the visual representation capability and thus strengthening the feature extraction capability of the backbone network for the smoke branch.The CBAM attention mechanism is introduced at the connection between the backbone network and the feature fusion network,so that the features extracted by the model can make full use of the spatial information and channel information of the feature map and enable the model to extract the target features in a more targeted manner without stacking the number of parameters heavily.(3)A self-built smoking dataset is retrieved from the web and photographed in realistic scenes,and the dataset is expanded using algorithms such as Mosaic data enhancement and random affine transformation to increase the number of small targets in the dataset,improve the model’s ability to detect small targets in realistic scenes,and thus enhance the generalization of the model.The expanded dataset is used to train and validate the improved model,and the ablation experiments demonstrate that the improvements proposed in this paper are realistic and effective.Experimental results show that the improved model proposed in this paper has higher accuracy and recall,and the improved model improves by 2.4% and 1.5% in m AP@0.5 and m AP@0.5:0.95,respectively,and by 0.5% and 1.0% in Precision and Recall,respectively.