Design And Implemention Of Pedestrian Tracking System In Intellingent Portraint Phtography

Intelligent portrait photography is an important component of the future photography ecosystem,which addresses the high demand for user skills in traditional photography by automatically capturing,recognizing,and tracking human subjects.In this technology,target tracking is the core algorithm that predicts the position,scale,and presence of a given target object in subsequent frames.Subsequently,intelligent photography devices control machines to track human movement or rotation based on the results of the target tracking algorithm.Intelligent photography devices have limited storage and computing capabilities on their onboard low-power platforms,making it difficult to apply state-of-the-art target tracking algorithms in practical engineering.This paper takes this as a starting point and designs a long-term target tracking algorithm that meets the precision and speed requirements of intelligent photography.We also construct a pedestrian tracking system and deploy it on a mobile photography robot.The main work of this paper is as follows:Considering that the existing long-term tracker cannot achieve the balance of speed and accuracy on the low-power platform,this paper designs Long-Term Lightweight Siamese Network(LT-LightSiam)based on the ELGLT(Effective Local and Global Search for Fast Long-Term Tracking).In the part of local tracking module,this paper designs LightSiam based on SiamFC++algorithm.1)In the local tracking section of LightSiam,in order to further reduce the complexity and computational complexity of the model,FBNet(Facebook Network)optimized specifically for mobile devices is used as the feature extraction network.Secondly,for the existing lightweight long-term tracking algorithms that cannot effectively handle target deformation and occlusion,and for the problem of tracking result boxes often drifting to the background or local areas of the target itself,this project introduces pixel level cross correlation as feature fusion to preserve more local information when designing templates and feature fusion of search areas,On the other hand,by combining search image features and related feature maps to enhance feature representativeness,thereby improving tracking accuracy.2)Next,this topic combines the advantages of two types of validation strategies to reconstruct the target state judgment strategy of the ELGLT framework,achieving dual validation and effectively improving the judgment of the target state.3)Finally,in order to verify the performance of LT-LightSiam,experiments were conducted on the long-term tracking dataset VOT2018LT/VOT2021-LT,and the results showed that the proposed method has significant improvements compared to the benchmark.This project deploys the above tracking algorithm on the mobile photography robot,and designs and implements object detection,object tracking,and motion control functions.The system is tested in real scenarios.