| As an important basic subject in the field of computer vision,object detection technology has been rapidly developed and widely applied in recent years with the explosion of deep learning and artificial intelligence technology.At present,the mainstream object detection technology based on deep learning has made great progress in all aspects of performance compared with the traditional object detection method,but there are still some problems in practical application,such as weak robustness to complex scenes,detection effect is greatly affected by the environment,and small targets are difficult to detect.In recent years,visual attention mechanism has developed rapidly,and it has been widely used in classification and semantic segmentation tasks.The integration of attention mechanism into deep convolutional neural network can significantly improve network performance,but there is no proposed attention model for object detection tasks.In view of the above problems,this paper studies the object detection method integrating the attention mechanism.By optimizing and introducing the attention module into the object detection model,the model detection accuracy and robustness are improved.The main research contents and innovative achievements of this paper are as follows:(1)To solve the problem of large number of parameters in CBAM model,a lightweight attention model(EAM)was designed.One-dimensional convolution and dilated convolution were introduced respectively to optimize the channel and spatial attention structure of the model.Compared with CBAM,the parameter complexity of the EAM proposed in this paper was significantly reduced.(2)Aiming at the problem of insufficient performance of existing target detection models,a YOLOv4 target detection model integrating attention mechanism was proposed.Based on the analysis of YOLOv4 network structure and application requirements of object detection technology,the improved design principle of YOLOv4 fusion attention module is proposed.After YOLOv4 fusion the EAM proposed in this paper,the mAP index is improved by 3.48 percentage points on VOC2012 data set.(3)The data set of abnormal behavior in examination room is constructed.Collect video stream data from multiple scenes and angles,cut,clean and mark the video stream,and standardize the abnormal behavior data set in the examination room according to the organization structure of VOC data set.Meanwhile,expand the training set samples by means of data enhancement to enrich the characteristics of diversity and alleviate the problem of sample imbalance in the training set.After data enhancement,a total of 191,582 samples were collected from all categories.(4)The identification model and algorithm flow of abnormal behavior in examination room are designed.According to the statistical results of abnormal behavior data sample distribution in the examination room,the adaptive optimization of YOLOv4 object detection network structure integrating the EAM was carried out,and the improved YOLOv4 detection model adapted to the abnormal behavior recognition scene in the examination room was proposed,and the algorithm flow of abnormal behavior recognition in the examination room was designed.MAP index decreased by 0.24 percentage points and FPS index increased by33% in the scene of abnormal behavior identification in examination room by improved network. |
