Visual Object Tracking Algorithm Based On Particle Selection With Deep Visual-semantic Information

The visual object tracking technology mainly uses the time continuity and the space continuity of visual objects in the video to track the object.The visual object tracking technology has been broadly applied in national defense,military affairs and security of residents etc.In addition,it involves the computer vision and pattern recognition technologies,which has important theoretical research value.However,the object and background in videos are changing with time.Besides,they are affected by a variety of factors such as illumination change,background clutter,object blur,and so on,which bring great challenges to the visual object tracking technology.In recent years,the deep learning networks trained on massive data has more formidable superiority in visual object characteristic extraction and representation compared to the general manual characteristic.It is suitable for the application of object tracking.High layer deep features are rich in semantic and category information,but with the number of feature layers increases,the spatial resolution of features becomes lower,which makes it impossible to accurately locate the object position.In this paper,according to the time and space continuity of the object in videos,candidate particles are obtained according to the position of the object in previous frame by particle filtering.Then,the particles selected based on the deep features are more relevant to the object.Furthermore,the related particles are precisely positioned according to the target detection algorithm or sparse representation.The research outputs obtained in this paper as follows:1)We propose a visual object tracking algorithm based on robust candidate particle selection with the combination of deep vision feature and object detection algorithm.Faster-RCNN extracts and generates feature vectors of candidate region based on feature maps,and uses them for accurate regression of candidate region.In this paper,sufficient candidate particles are generated by particle filtering algorithm,and then a small number of candidate particles more similar to the object are selected by deep features generated by the deep learning model for final verification using Faster-RCNN.The experimental results show that this method can not only improve the accuracy but also ensure the robustness of the algorithm.2)We proposed an algorithm to filtrate candidate particles based on the full convolutional deep features,and select the particles with the highest reliability mostly related to the object and the local sparse representation template with color information for minimum reconstruction error.The particle with the smallest reconstruction error is the predicted object.The algorithm firstly selects the visual features extracted from the deep model using a heatmap model and selects the feature map with the largest response value in the object area.Then,the object appearance is reconstructed combining with sparse representation,and the selected candidate particles are matching with the object template with its similarity degree to predict the object.The algorithm in this paper combines the deep learning model with the traditional sparse representation,which can not only extracts the high-level semantic information of the object,but also use the discriminative feature of the object appearance feature,which guarantees the robustness of the tracking algorithm.