Key Techniques Of Dynamic Object Detection And Tracking In Ground And Air Environments

The rapid development of artificial intelligence technology has brought a more convenient and efficient way of life to human society.In recent years,benefiting from the rich expressive power of depth models,computer vision has gained wide application in multiple fields,Object detection and tracking is two important parts of it.its main task is to find the target to be measured from the image,output its category and position,then analyze the motion pattern of the target to achieve stable and effective tracking.The frequent use of thermal weapons in complex ground and air environments generates smoke,floating soil and flames,which bring a lot of illumination variation and image blurring,and the targets are more easily obscured by obstacles;in addition,the targets in images are generally small and there are a lot of scale changes and shape changes.All these difficulties lead to a reduction in detection and tracking accuracy.It is a difficult task to detect and track dynamic objects accurately in real time.In national defense construction,dynamic real-time tracking of enemy targets can realize visual positioning and provide guidance for defense and striking target.Making detection and timely warning of abnormal state of our units is beneficial to our timely defense,or battlefield rescue.In view of the difficulties unique to the complex ground and air environment in object detection and tracking,some improvement measures are put forward to adapt to the complex ground and air environment.(1)For the object tracking in complex ground and air environment,the Deep SORT tracking method based on ResNet-18 is studied to solve the problem of low tracking accuracy in the presence of a large quantity of small targets and illumination variation by deepening the appearance extraction effect.The method reduces the ID Switch phenomenon in the process of small object tracking and improves the tracking stability.(2)A ResNet-based CBAM-YOLOv5 detection model is designed for accurate detection of small targets in blurred images under complex ground and space environments.The YOLOv5 structure is transformed by using CBAM to improve the detection accuracy of the network for small targets and fuzzy images,and to optimize the multi-scale detection effect.In addition,the ResNet-34 network is introduced to reclassify the targets that may be misclassified by the CBAM-YOLOv5 network and optimize the detection results.The experimental results show that the proposed detection algorithm outperforms the commonly used detectors such as YOLOv5 network in meeting the premise of real-time detection and improves the detection accuracy and recall rate for small targets and fuzzy images.(3)To address the problem of anomaly detection of dynamic targets in complex ground and air environments,the ResNet-based CBAM-YOLOv5 smoke and fire detection model is deployed on Ascend 310 using MindX SDK.The accuracy of the model for smoke and fire target detection is tested under various scenarios,and the performance and false detection rate of the proposed detection model are compared with other commonly used smoke and fire detectors.The experimental results show that the proposed smoke and fire detection model can accurately detect smoke and fire targets in multiple scenarios,and the accuracy and false detection rate are better than those of conventional detectors.