Research And Application Of Indoor Multi-target Detection Based On Deep Learning

With the arrival of industry 4.0 era and the formulation of made in China 2025,China's industry has gradually developed into the field of intelligent manufacturing,transforming from a manufacturing power to a manufacturing power.In the traditional machinery manufacturing industry,materials need to be handled manually,and a large number of handling equipment is needed in the process of material handling.In the process of using,it is common to place the handling equipment at will.It takes more time to find and check the equipment each time,which is not conducive to improving the production efficiency.In this paper,the indoor multi-target detection based on deep learning is applied in the scene of factory indoor target detection.The target equipment is identified and located by the camera,which is convenient for the porter to quickly find the required equipment.It is of great significance to improve the production efficiency,reduce the idle rate of equipment and improve the factory intelligence.Aiming at the problems of illumination,occlusion and deformation in indoor scene target detection,this paper focuses on the optimization of indoor scene multi-target detection network,and describes the application of deep learning in indoor industrial scene multi-target detection,taking equipment detection as an example.The main contents of this paper are as follows:1.Test and Research on the algorithm of Yolo v3Yolo v3 algorithm is one of the most widely used target detection algorithms,which has a good performance in speed and accuracy.After the detailed introduction of the Yolo v3 algorithm,this paper uses the complex scene training set to train the Yolo v3 network model.After the training,this paper compares it with the target detection algorithm SSD based on linear regression and the fast RCNN algorithmbased on the candidate region recommendation algorithm in the horizontal comparison test.As shown in the test result,the detection accuracy of Yolo v3 algorithm depending on the residual depth network is better than SSD.But due to the difference of detection mechanism,the detection performance of target detection based on candidate region through two-stage network is better than the algorithm of single-stage linear regression,and the detection accuracy of Yolo v3 is only 2.3%lower than fast RCNN.Based on these study,we can draw a conclusion: the difference of feature extraction network has a significant impact on the model accuracy.2.Network optimization based on Yolo v3 algorithmThrough the analysis of Yolo v3,it is concluded that there are several defects in the detection of indoor targets,one is the weak detection ability on occluded targets,the other is the poor generalization ability of the model.Based on the above two deficiencies,this paper uses the feature extraction mechanism of multi receptive field to enhance the feature extraction ability of Yolo v3.Through the experiment,the test results of the experiment show that more abundant feature information is helpful to the improvement of the network accuracy and the enhancement of the generalization ability.Through the comparative test of the model,the average accuracy of the Yolo v3 network optimized by the multi receptive field mechanism is improved by 3.3%compared with that before the optimization.For the detection of some complex targets,the accuracy is improved by more than 6%.The fitting degree between the prediction box and the annotation box is also improved by 78.7%(average IOU),which is also improved by 3.5% compared with that before the optimization.The feature extraction mechanism of multi receptive field is of great help to improve the performance of network detection,and the practicality and overall robustness of the optimized network model after the system is carried on is also improved significantly.The optimized network is suitable for complex scene target detection.3.Experiment and analysis of target detection and location in factory indoor sceneThe optimized Yolo v3 network model is applied to the target detection of indoor scene in the factory,and the data set of common tools is constructed.The optimized network structure is trained and tested.Through comparative test,the optimized network model has higher accuracy in industrial scene detection.This paperintroduces two kinds of target location algorithms,and tests the two kinds of target location algorithms.Through the comparison test,the average error of target location algorithm based on area estimation is 1.32 meters,which can meet the calculation of target equipment position.4.Design and implementation of factory indoor target detection systemPTZ camera equipped with target detection and target positioning algorithm is used to detect and calculate the location of the indoor scene target in the factory,and the detection image and detection information are transmitted to the client.Through the analysis and research of the system function,the factory indoor target detection system is designed.The client can display the detection screen,search tools and view the target position.After testing,the system has higher detection accuracy and can meet the detection requirements.Through the test of the factory indoor scene data set,the indoor scene multi-target detection method based on deep learning has certain effectiveness,and realizes the indoor target detection task in the factory.