| Currently,mostly drowning detections in swimming pools are manual supervisory based on the camera information,which is not only labor-intensive but also causes miss detection or false detection due to the uncertainty of human activities.Just a few drowning detections are automatic monitoring based on computer vision,among which the representative algorithm is background subtraction.However,the detection speed of this algorithm is slow,only about 5 fps.Besides,detection accuracy is low,only about 60%.These problems make the drowning cases can not be timely or accurately detected,which results in drowning people can not be rescued in time.To solve the above problems,this paper applies deep learning methods in automatic drowning detection based on computer vision and obtains several deep learning models for drowning detection.The specific work contents are as follows:(1)For the problems of slow detection speed and low detection accuracy in the traditional background subtraction method,this paper selects the representative detection model Faster R-CNN as the initial model,and then get the final model by training on the self-made COCO drowning data set.Finally,a fast and accurate drowning detection model under complex background is obtained.The experimental results show that compared with the traditional background subtraction method,the detection speed of the proposed drowning detection model is 12.1 fps,which is about 1.4 times of the traditional background subtraction method.The detection accuracy of the model is 82.31%,which is improved by 35%.The proposed drowning detection model is much better than the traditional background subtraction method in detection speed and accuracy aspects.Our method can get faster and more accurate drowning detection in the complex swimming pool.(2)For the problems of low detection accuracy and vulnerability to interference in the Faster R-CNN model,this paper firstly uses the Mask R-CNN model as the initial model,and then trains the model based on the self-made COCO drowning dataset to get the final model.Finally,an anti-interference drowning detection model under complex background is proposed.Experimental results show that,compared with the Faster R-CNN model,the detection accuracy of the proposed drowning detection model based on the Mask R-CNN model is 90.51%,which is improved by 9%～10%.In addition,the proposed drowning detection model can distinguish the swimmer from the complex background of the swimming pool,which reduces the background interference of the drowning detection.(3)For the problems of miss detection or false detection in the model Mask R-CNN,as well as the missing anchor box information in the middle of the threshold when the model marking positive and negative samples according to the IOU threshold of the real box and the anchor box,the AS-SAG-Mask R-CNN based on an adaptive selection of positive and negative anchors and attention mechanism is proposed in this paper.Based on the results of(1)and(2): compared with the model Faster R-CNN,the Mask R-CNN not only has higher detection accuracy,but also separates the swimmer from the pool background and reduces background interference.Therefore,the drowning detection model is based on Mask R-CNN,and the model was obtained by adopting automatic selection of positive and negative samples of the anchor frame and introducing the spatial attention mechanism on the mask branch.Since it improves the initial Mask R-CNN drowning detection model and we train the model based on the sequence of video sequence frames,the final drowning detection model based on the automatic selection of positive and negative anchor frames and attention mechanism has learned some drowning dynamic characteristics.The experimental results show that compared with the model Mask R-CNN,the detection accuracy of AS-SAG-Mask R-CNN model reaches 94.83% while maintains a high detection speed,which is improved by 4%～5%.What is more,it can effectively avoid missed detections and false detections in drowning detection. |
PDF Full Text Request