Research On Multiscale Object Detection At Construction Site

With the rapid development of urbanization and the expansion of building scale year by year,building construction safety issues have also been paid more and more attention.How to improve the efficiency of building construction safety management has become an urgent problem to be solved.With the rise of the artificial intelligence industry,the rapid development of computer vision technology,especially deep learning,has brought solutions to this problem.Therefore,effective use of deep learning methods to achieve safety monitoring of construction sites is of great significance.In order to realize the safety monitoring of large scene construction sites,it is necessary to carry out target detection and instance segmentation for various construction machinery and personnel.However,the current mainstream deep learning target detection algorithm has the problems of low detection accuracy of small targets and poor boundary segmentation effect,it is easy to cause construction safety hazards if they are directly applied to construction safety detection in large scenes.Therefore,multi-scale target detection is still difficult and challenging.In order to improve the effectiveness of intelligent safety monitoring,this paper mainly uses deep learning algorithms to carry out multi-scale target detection and case segmentation research on excavators,cranes,and construction personnel in construction sites.In response to the above issues,two improvement schemes are proposed combining data enhancement and network improvement.The specific work is as follows:(1)An enhanced multiscale target detection algorithm was developed to solve the problem of poor detection performance due to scale changes in construction sites.First of all,automatic expansion of scale aware data is defined to learn data expansion strategies.Then,in order to reduce the information loss caused by the reduction in the number of channels when using a feature pyramid network,we proposed a method based on subpixel convolution to perform channel enhancement and upsampling,and added a bottom-up path to enhance the entire feature hierarchy using accurate positioning information in the low-level high-resolution image.Experimental results show that the algorithm achieves good target detection accuracy on construction site(MOCS)dataset and MS COCO dataset,especially for small targets.(2)An improved Mask R-CNN model based on global context channel attention(GCCA)mechanism and multi-stage refinement mask is proposed.First,this paper gradually fuses finegrained features in the mask head in a multi-stage manner to refine high-quality masks.Secondly,in order to better integrate fine-grained features,the GCCA attention mechanism is constructed,which aggregates global features through a simplified global context module,and uses one-dimensional convolution to achieve local channel interaction without dimensionality reduction.Experimental results show that the improved algorithm achieves higher accuracy in detection and segmentation.