Vision Based Detection And Tracking Mehtod For Swimmers

The technologies of object detection and object tracking are research hotpots in the field of computer vision,and play a key role in many fields,such as autonomous driving,intelligent monitoring and sports competition,which bring great convenience to people’s life.With the continuous development of science and technology,various advanced detection and tracking algorithms have been proposed.However,in the actual application scenarios,they are still challenged by scale changes,background interference and other factors,and there are still various problems and difficulties in object detection and tracking that need to be studied and solved.This paper mainly studies the detection and tracking method of swimmer based on vision.The specific research contents are as follows:Firstly,aiming at the multi-scale object detection problem,this paper proposed a feature fusion.The network enhances the feature map by fusing different layer features,so as to improve the performance of multi-scale detection.By adding a horizontal feature enhancement module in the feature fusion network,the problem of deep feature loss in the fusion process is solved.In addition,we added the shallow feature fusion path.At the same time,by designing the shallow feature fusion module,we used attention mechanism to replace the feature fusion method of adding elements directly.Experiments show that the feature fusion network can effectively improve the performance of network detection.Secondly,by fusing FHOG and color features,the two features are synthesized to obtain a complete representation of the object.Then,we improved model update strategy for model pollution problems caused by fixed learning rate updates of trackers.And we proposed the coefficients PCF and HIST from the point of view of the concentration of response graph and the similarity of objects,which can indirectly reflect the tracking effect.After that,by judging the tracking effect of each frame image in real time,it dynamically determines whether the model is updated,which effectively reduces the impact of poor results on the model.Experiments show that the model updating strategy can effectively make up for the defects of the fixed learning rate updating mechanism and improve the tracking performance of the algorithm.Finally,aiming at the tracking problem of the actual swimming process,this paper proposes a comprehensive detection and tracking algorithm.Identify the problem for the initial object,and we use the object detection algorithm output result to complete the initialization of the tracker.After that,the test results are comprehensively filtered by the confidence degree and area size of the prediction box,which solves the problem that the confidence level does not fully reflect the reliability of the results.In addition,in view of the problem of drift in the tracker for a long time tracking and the accuracy decrease,we realized the long-term object tracking by increasing the abnormal situation judgment and object re-detection link.At the same time,the object tracking position is corrected by the Kalman filtering method.Finally,the effectiveness of the prosed method is verified by experiments.