Font Size: a A A

Research On Detection And Tracking Of Small Balls In Rhythmic Gymnastics

Posted on:2022-08-07Degree:DoctorType:Dissertation
Country:ChinaCandidate:Shambel Ferede AssemahagnFull Text:PDF
GTID:1487306602993929Subject:Intelligent information processing
Abstract/Summary:PDF Full Text Request
In recent years,detection and tracking of objects in videos have got a great attention from several computer vision researchers due to their different applications,such as traffic control,robotics,self-driving vehicles and surveillance.In this dissertation,research on detection and tracking of small balls in rhythmic gymnastics is proposed,and its main goal is to detect and track the movement of small balls in a video sequence.This study is helpful for the coaches to grasp the accuracy of actions of the players and for the growth of intelligent physical education enlarge.Furthermore,the contributions of this dissertation are a robust small ball detector named channel feature enhanced detector(CFED)is proposed,channel feature enhanced module(CFEM)is proposed to increase the discriminability of target feature map,BALL dataset is organized for the experiment,proposed small ball tracker,and proposed to incorporate temporal and direction information to track small balls effectively.The core challenge in detection and tracking of small balls in rhythmic gymnastics is that the features of small balls is very vague due to deformation,motion blur,shadowing and the color and shading of background similar with the balls.To overcome these challenges,detection-based tracking method is applied.First,to detect small balls efficiently CFED is proposed.CFED has improved SSD to make it appropriate for small ball detection and CFEM is added for feature prediction in order to increase the discriminability of the target features.Squeeze-and-Excitation block is applied for CFEM by exploiting the inter-channel association of features for the task on the proposed BALL dataset and the module computes the importance of each convolution kernel of the feature layer.We used global average pooling layer to obtain the durable response value of each channel feature map,L2 function is used to normalize the amplitude of the vector and scale layer is used to get the final output.The proposed BALL dataset is organized for training and evaluation.The proposed network is applied based on SSD build on Caffe framework and VGG16 architecture.The small ball detection experiment result shows that the proposed detector,CFED,achieves a m AP of 90.2%which is more accurate and it takes an average speed of 11.4 ms per image.Further ablation study was executed to validate the success of various components in CFED with the same training iteration,batch size and input size.The test results compared with and without CFEM.The accuracy improves from 85.0%to 88.1%with CFEM and the test results confirmed the effectiveness of CFEM.The effectiveness of the batch normalization(BN)is also evaluated by comparing the test results with and without the BN layers.With BN the accuracy increases from 85.0%to 88.9%and it shows that the BN is a valid training method for arbitrary initialization.Global average pooling technique is used to obtain the weight of each channel in the proposed network and another experiment has done with the global maximum pooling approach.The test result shows that global average pooling is more efficient than global maximum pooling.Second,small ball tracking with trajectory prediction is proposed to track small balls when the athletes playing with them at sports field.Due to its significant detection quality on the tracking performance,CFED is selected as a detector for the tracking process.The proposed tracking technique is based on Deep SORT and improved it for the tracking of small balls.Due to lack of appearance information for small balls,deep SORT can't track them well.To solve this problem,the designed algorithm incorporates motion,frames storage and directional information to indicate the trajectory.The last various frames that have matched detections are retained for re-identifying when the small ball reappears and this can solve the problem of ID failure,and also direction judgment is added to solve object overlap problem.Trajectory is used to indicate the motion and direction of small balls in the small ball tracking process.A set of trajectories are engendered from the set of small balls to locate the small balls using its trajectory and to predict the locations of the missed balls.The small balls tracking experiment implemented on a collection of proposed BALL datasets.The experimental result of the small ball tracking performance is evaluated in comparison with deep SORT based on the metrics evaluation.The experimental result displays that the proposed improvement effectively decreases the number of identity switches from 137 to111,track fragmentation from 496 to 420,false detections from 891 to 809 and similarly the number of missed detections reduces from 1201 to 725.Thus,there is significant increment in the number of small ball tracking accuracy from 49.5 to 62.7 and precision from 66.2 to67.3,and also the speed of the proposed method is greatly improved from 19.17Hz to156.34Hz.The general result shows that the proposed methods effectively detected and tracked small balls in rhythmic gymnastics.
Keywords/Search Tags:object detection, object tracking, small object detection, small object tracking, intelligent physical education
PDF Full Text Request
Related items