Research On Object Tracking Under Complex Scene Based On Convolutional Neural Networks

Visual object tracking is an important branch of computer vision,which uses the contextual information of video or image sequences to model the appearance and motion of the target object,thereby predicting the target locations.However,the unknown target and track scene,the limited provided initial state of the target,and the changing appearance of the target over time,make the general single object tracking problem extremely challenging.This paper focuses on the research of the key technology of visual object tracking based on convolutional neural network,including the rotated scenes,interference of the similar objects,occlusion,etc.,to improve the robustness and accuracy of visual object tracking in complex scenes of visible light video.The main contents of this paper include:To address the problem that traditional tracking algorithms only output bounding boxes without rotation angles,we present an object tracking algorithm based on siamese convolutional neural network with multi-scale and multi-rotation detection,to accurately detect target boundary and estimate the rotation status of the target.The proposed method uses the siamese convolutional network to directly learn the similarity function between the search region and the template.To estimate the scale and rotation state of the target,we utilize data augmentation by transforming the search region with different scales and different rotation angles.Based on the past locations of target,a motion model is proposed to constraint the search region in the next frame.The experiments show that the proposed algorithm with a simple structure presents the accurate boundary positioning capabilities and fast processing speed,especially under the rotated scenes.To address the drifting problem caused by the interference of the similar objects nearby,we propose a discriminant network based on the discrete correlation filter for object tracking.By constructing a shallow siamese convolutional network with a special discrete correlation filter layer,the discriminant network is able to update online to learn the difference between similar objects and the target.By adding this discriminant network to a normal siamese network,the normal siamese network searches coarsely and the discriminant network searches finely,thereby discriminating the similar objects and the target.In order to construct a lightweight network and remain performance,the vector convolution kernel is introduced to compress the parameters of the network.The experimental results show that the algorithm performs better than the normal tracking algorithm based on the siamese network in the overall performance and under the interference of similar objects in the background.To address the occlusion problem,we propose a novel tracking algorithm using the Spatio-temporal network for trajectory prediction.This algorithm learns the past information of the target and uses trajectory prediction to solve the problem of tracking the occluded target under the camera moving scene.This method consists of three networks,including a Siamese based background-motion network that matches the global background motion between adjacent frames for compensating the camera motion,an appearance inference based tracking network,which can be any existing tracker based on deep learning,and a trajectory prediction network to predict the future locations of the target based on the background motion vector,target's past trajectory and few previous frames,meanwhile,a sub-classified network estimating the confidence of the predicted trajectory.The experimental results show that the algorithm is able to assist the existing tracker to handle occlusion in the complex scene of camera movement and the interference of similar objects in the background.To address the above trajectory prediction algorithm with the problems of training without end-to-end and the over-fitting confidence,we propose a trajectory prediction and tracking unified model with confidence calibration.This method also aims to address occlusion problem.The difference with the above trajectory prediction algorithm is that,this method integrates the background motion network,tracking network,trajectory prediction network and assessment network into a unified model,which can be trained end-to-end.This unified model takes in a few previous frames,and outputs a tracking result in the current frame and the trajectory prediction locations in the current frame and future frames,with their confidence scores.The confidence score is estimated by an assessment network based on classification.Due to the general classification network trends to output higher confidence,a calibration model based on temperature scaling is introduced to calibrate the trajectory confidence score for matching the true probability.The experimental results show that the proposed method is able to handle occlusion more effectively compared with the above algorithm.In summary,the four deep learning-based algorithms proposed in this paper effectively solve the rotation,similar object nearby,and occlusion problems in visual object tracking under complex scenes.