Research On Target Intelligent Search And Tracking Technology For Edge Applications

Nowadays,with the development of artificial intelligence and computer vision,the specific target search and tracking technology based on real-time video data stream has gradually evolved into its core task,which has been widely used in aircraft guidance,key area monitoring and human-computer interaction scenarios.The target search and tracking method based on deep learning has good search and tracking effect,while the complexity of the deep learning model makes it have higher requirements for the computing platform.How to transplant and deploy the deep learning target tracking model on the portable edge to solve the practical problem of intelligent technology has become a hot spot in the field of computer vision and artificial intelligence.In this thesis,the specific target under real-time video data was regarded as the object of identification and tracking,and the application technology and method of target intelligent search and tracking under the background of edge application was studied,and a set of target intelligent search and tracking system for edge application was also designed.The main research contents include the following aspects:(1)Based on the YOLOv3 model,we proposed a lightweight target detection network suitable for edge computing platforms.Multiple secondary convolution kernels were extracted from the Res Net-50 network using the secondary auxiliary convolution technology firstly;Then the large receptive field and small computational complexity of the secondary auxiliary convolution were utilized,and multiple secondary auxiliary convolution kernels were connected to reconstruct the traditional convolution kernel of the backbone network,so as to reduce the network computing consumption on the premise of maintaining high detection accuracy of the target detection network;Finally,a comparative experiment was performed by comparing the public target detection data set with the edge computing end and the original detection backbone network model.The results showed that the accuracy of the lightweight detection model(TOP-1)was only reduced by 0.12%,maintaining the detection accuracy of the original model,while the memory footprint was reduced by 22% of the original model compared with the number of floating point operations per second,and the average load value for edge-use applications decreased by51.2%.(2)Infrared picture image and sound source multi-modal features were used to assist target detection tasks in different application scenarios,and different edge devices such as PTZ cameras and far-field microphones were used to realize the intelligent target search.The resolution of the imaging image in the large-scale outdoor environment was too high,resulting in the low accuracy of target detection.The infrared significant area was used for dual light image registration with the corresponding visible light area,and the target intelligent search in the high-resolution image scene can be realized with the help of three-dimensional zoom pan tilt.On the other hand,due to the lack of target correlation recognition in the indoor environment,resulting in low hit rate of target detection task.Far-field microphones and Baidu cloud speech recognition were used to pick up and identify far-field sound sources respectively,also PTZ cameras were applied to improve the situation.Through the experimental verification in different environments,the multimodal feature assistance has been proved to effectively improve the detection efficiency and hit rate of the target detection task,realizing the intelligent search of the target.(3)Based on the previous work,cascade the Deep Sort target tracking network and the Re ID re-identification network to realize the target intelligent search and tracking for edge applications,and transplant and deploy it on the edge computing platform Xavier.Through the comprehensive performance comparison,the accuracy of the target search and tracking algorithm in this study increased by 0.8% over the MOT16 data set,the average center point error decreased by 5.685 pixels,IOU coverage increased by 8.504%,and it has a low CPU and memory occupancy rate during runtime at the edge.Overall,the study results can provide key technical support for the edge implementation of target intelligent search and tracking.