Research On Multiple Object Trajectory Fusion Based On Deep Learning And Multiple Hypotheses Tracking

Since the 1970 s,visual tracking algorithms have made great progress.And multi-target tracking is one of the hot topics.According to the classification of initial state,multi-targets tracking is mainly divided into two categories: Detection-Based Tracking(DBT)and Detection-Free Tracking(DFT).According to the processing mode,there are two types of multi-target tracking algorithms: Online tracking and Offline tracking.This thesis focus on multiple object trajectory fusion based on deep learning and multiple hypotheses tracking,which belongs to the category of Online tracking and DBT.The Mask R-CNN is used as the target detector to detect the position information and confidence of the video target in each frame,and extract the deep features of each detection frame.Then the PCA is used to reduce the dimension.On the data association step,we use Multiple Hypothesis Tracking(MHT)algorithm,which is the improvement of traditional MHT.In order to improve the tracking speed and ensure the tracking accuracy,it is necessary to increase the gated area of data association and add additional pruning.However,it is not enough to use motion information for data association.We use Multi-output Regularized Least Squares to conduct appearance model.The experimental data proves that the appearance model can add additional pruning and improve the tracking accuracy.The appearance model can also reduce the dependence of MHT on the parameter setting.So,the method of expanding the gated area can be used to improve tracking speed.The improvement of MHT reduces the amount of calculation and improves the tracking speed using the pruning method.This thesis using the multiple targets tracking method of Mask R-CNN + the improvement of MHT not only improves the tracking precision,also speeds up the tracking speed.The method has the very good robustness.A multiple target trajectory fusion algorithm based on deep learning and multi-hypothesis tracking is proposed.We extract the color feature,HOG feature and deep feature of the each bounding box.We use the three kinds of features to construct appearance model for getting three tracking trajectories.Then,we merge target tracking trajectories of the three different appearance model tracking algorithm and reach the optimization of tracking trajectory.It improves the performance of the multiple targets tracking algorithm.