| The 14 th Five-Year Plan of China emphasizes the importance of promoting high-quality development in its manufacturing industry and constructing a strong and sustainable manufacturing powerhouse.It is also pointed out that the upgrading and transformation of traditional industries should be promoted through the empowerment of industrial internet and intelligent manufacturing.However,the production and management of the traditional PCB industry still heavily rely on the proficiency and skill level of employees.Providing relevant guidance information quickly and accurately for assisting operators in production,inspection,and management is of great guidance value and practical application.Currently,most PCB companies still use paper-based work instructions,and only a few companies have realized informatization on PC,mobile terminal,and data screens.With the development of the "Industry 4.0" plan,some experts and scholars have combined AR technology with industrial environment to develop industrial AR application system,and whether the object tracking of industrial equipment can be done accurately in real time will directly determine the performance of AR system.Due to the varying poses and appearances of industrial equipment in images,as well as various factors such as occlusion and background interference,there are great challenges in applying object tracking algorithms in current application scenarios.To address the above issues and improve the accuracy of multi-object tracking algorithms in practical industrial applications,this study conducted relevant research based on the YOLOv7 and Deep Sort algorithms.The research work and contributions are as follows.(1)To address the low detection accuracy of detection-based multi-object tracking algorithms in the PCB industry due to the dense placement and overlapping of equipment,an improved YOLOv7 industrial equipment object detection model was proposed.Firstly,the CA attention mechanism was introduced into the backbone network to improve the network’s attention to the visible area features of PCB industrial equipment.Secondly,the SIo U was used to improve the detection model’s loss function,effectively enhancing the accuracy of object position regression.Finally,the Adaptive-NMS algorithm was introduced to adaptively adjust the threshold based on the object’s density to retain more accurate prediction boxes.Through experiments on a self-made PCB industrial dataset,the experimental results demonstrated that the improved object detection model can significantly improve the detection accuracy in environments with overlapping obstructions.(2)To address the issues of object occlusion and high model complexity that affect the real-time and accuracy of trackers,the feature extraction model of the multi-object tracking algorithm Deep Sort was improved.Firstly,the lightweight network Shuffle Net V2 was introduced into the feature extraction model of the Deep Sort multi-object tracking algorithm and retrained to reduce the model’s parameter volume and improve its inference speed.Combining with the YOLOv7 detection model,more accurate boundary boxes were obtained,and the tracking accuracy was improved.The improved multi-object tracking model was experimentally evaluated on a self-made PCB industrial dataset,and the results demonstrated that the algorithm can achieve good tracking accuracy and real-time performance in crowded industrial equipment placement environments.(3)To verify the applicability of the improved multi-target tracking algorithm in AR application systems,a prototype AR-based operational guidance system was built and implemented.The prototype system was combined with an improved algorithm based on YOLOv7 and Deep Sort,and multi-object tracking tests of PCB industrial equipment were conducted using Hololens2 devices as the carrier.According to the system’s operational results,the prototype system combined with Hololens2 devices can achieve accurate detection and real-time tracking of industrial equipment in PCB production workshops. |
