Research On Target Tracking And Fall Detection Algorithms Based On Deep Learning

As one of the research hotspots in the field of computer vision,target tracking has a wide range of applications,such as autonomous driving and intelligent monitoring.However,since the target may experience scale change,geometric deformation,occlusion,and background clutter during the tracking process,the tracking difficulty of the algorithm is increased.This paper proposes a new target tracking algorithm to address the above challenges.In addition,fall detection has great application prospects.Most traditional fall detection products use wearable devices,which have a greater impact on the daily activities of the elderly and are relatively expensive.Therefore,this paper proposes a computer vision-based fall detection algorithm.This article first conducts an in-depth study of the target tracking algorithm based on correlation filtering,and discusses how to improve the performance of the tracker in a complex environment.In addition,inspired by the target tracking algorithm,the IOU tracking matching algorithm is added to the classic AlphaPose extracted skeletal feature network.The accuracy of this tracking algorithm directly affects the success rate of fall recognition.The contributions of this article are as follows:Through research on target tracking,this paper proposes a target tracking algorithm based on rotation adaptive multi-feature multi-template fusion(RA-MFMT).First,given the situation that some serious background interference information causes the tracker template pollution,a multitemplate learning model is designed.Through the construction of the global template,the decision template,and the correction template with complementary characteristics,the tracker can improve the anti-interference ability in the complicated scene.Secondly,because there are a large number of interference sources in the tracking process,the algorithm uses shallow color features as visual supplementary information to adaptively fuse the multi-layer depth features of VGGNet-19,which can improve the tracker's ability to discriminate the appearance of the target.Meanwhile,an adaptive target rotation angle estimation strategy is proposed to solve the problem of tracker performance degradation caused by target rotation.This strategy uses an improved tracking result confidence to estimate the target rotation angle.And on the OTB-2013,OTB-2015 and LaSOT data sets,the performance of the RA-MFMT proposed in this article and a variety of mainstream algorithms are compared.In addition,through the research of target tracking algorithm,this paper proposes a Mobialphapose fall detection algorithm,which includes a video preprocessing module,a feature extraction module,and a fall decision module.First,the video preprocessing module uses the Gaussian mixture model for foreground detection.Then,we use a contrast limited adaptive histogram equalization and Gaussian blur strategy to enhance the foreground area and weaken the background area,respectively.Moreover,since the skeleton sequence estimated by the traditional AlphaPose network only extracts position features and ignores the problem of temporal features,this paper improves the network by adding IOU matching method between the pedestrian detection module and the single-person bone detection module.Further,the fall judgment module is composed of threshold judgment and MobileNet V2 network judgment.Among them,the threshold judgment method is first used to eliminate the obvious fall behaviors.For the suspected fall behaviors,the lightweight MobileNet classifier is adopted to make second discrimination.Finally,the proposed Mobi-alphapose algorithm,the smallest bounding rectangle method,and Chua's method are performed on the self-built dataset.Compared with the two comparison algorithms,Mobi-AlphaPose has a high fall recognition success rate and a low false detection rate.