Research On Construction Machinery Object Detection Based On Faster R-CNN

With the continuous development of science and technology,video monitoring technology has been widely used in the construction site.Through video monitoring images,we can dynamically understand the construction progress,order and safety.However,traditional video surveillance only provides video surveillance images,which still needs a lot of manual troubleshooting and analysis,and can not meet the needs of realtime monitoring management.Therefore,it is the general trend to realize intelligent video surveillance image analysis system.And one of the most important links is the object detection of video surveillance images on the construction site.The main research work of this thesis is the research and application of object detection of construction machinery,crane,excavator and hoist on the construction site.There are many difficulties in the object detection of construction machinery:the object color and shape of construction machinery are different,and the object size is different;The construction scene is extremely complex,and there is often mutual occlusion between objects;and the image quality of video shooting is easily affected by uncontrollable factors such as weather.In order to realize the accurate detection of the object,this thesis focuses on the above problems,and the main research contents are as follows:Firstly,this thesis propose a construction machinery object detection method based on multi-scale feature fusion.(MSFF-FRCNN).This method is based on the idea of Faster R-CNN and has the following characteristics:(1)the K-means clustering algorithm is used to guide the setting of the aspect ratio of the prior anchor box,making it more suitable for the size of the object studied in this thesis;(2)use generalized intersection over union(GIOU)as the bounding box regression loss;(3)use ResNet50 as the backbone feature extraction network,combined with the constructed feature path aggregation network(PFPN)to fuse the object multi-scale features.This method can also achieve good detection results for objects with large differences in size.Secondly,this thesis proposes a method for object detection of construction machinery based on attention and feature fusion(AT-FFRCNN).This method is based on MSFF-FRCNN,and introduces an attention module CBAM(AT-RPN)in the region proposal network,so that the network can pay more attention to the feature channels and spatial positions related to the object,and then combines the attention mechanism in the fully connected layer for enhancement(AT-RPN),using the surrounding proposal boxes to update the feature vector of the object proposal box,which improves the accuracy of detecting occluded objects and unobvious objects.Thirdly,this thesis proposes an object detection method based on object region decomposition and combination(RDC-FRCNN).The method takes the detection of crane objects as an example,decomposes the crane object into two parts,the beam(TowerBeam)and the main body(TowerBody)as two separate object categories,and finally,the two regional object boxes are combined,and the smallest rectangular box containing these two object boxes is used as the final object prediction box,which reduce the classification calculation and regression calculation of the non-characteristic part(background)of the crane by the network and improve the speed of object detection.Finally,some experiments are carried out in this thesis.The experimental results show that the proposed method can complete the detection task of construction machinery objects well.