Research And Application Of Object Tracking Based On Siamese Network

Target tracking is one of the more difficult research topics in the field of computer vision.The purpose of this task is to track a target in the video and find out where the target appears in each video frame.The application scenarios of this technology are also relatively extensive,such as solving pedestrian tracking in autonomous driving tasks,tracking certain specific target objects in security monitoring systems,and so on.With the development of deep learning technology,more and more methods based on deep neural networks have been proposed and used to solve computer vision problems.The mainstream method of target tracking based on deep learning is the Siamese neural networks.The main idea of the algorithm is: The video data is split into several frames of image data and operations are performed sequentially.Use a lightweight convolutional neural network to extract the features of template image of the video and the search space images;perform cross-correlation calculation on the obtained feature vectors to obtain the similarity between the feature at different positions in the search image and the target feature;The feature with the largest cross-correlation response value is determined as the target feature,and the position of this feature vector is the position where the target appears in the current video frame.The method proposed in this paper is based on the Siamese neural networks and region proposal network.The innovations are as follows:Aiming at the imbalance of positive and negative samples in the region proposal structure,this paper proposes the Overlap Sampler sampling strategy to enable the network to screen positive and negative samples that are worth learning.Overlap Sampler calculates the IOU between all proposals and selects those negative samples that are close to the positive sample of target to participate in network training.For positive samples,the Overlap Sampler sampling strategy weights the best matched proposal of the target to make the network more focused on learning the positive samples that are closest to the target Ground-truth.The proposed method has improved the target tracking performance on the VOT2015 dataset.In addition,this paper proposes an adaptive feature enhancement module for low-quality videos to improve the accuracy of target tracking.We use the Artificial Bee Colony algorithm and the incomplete Beta functions in combination to fulfill requirements.The incomplete Beta function is used to perform image grayscale transformation.The Artificial Bee Colony algorithm is used to search the two parameters of the incomplete Beta function,so that we can obtain the corresponding optimal grayscale transformation function according to the quality of each frames.In addition,this paper proposes a video frame quality evaluation function thatcombines four image attributes,such as image edge response value,number of image edges,image entropy and image contrast to guide the learning of artificial bee colony algorithms.The adaptive feature enhancement module is computationally friendly and fast.It improves the accuracy of target tracking in low-quality videos without losing the speed of the algorithm,and achieves better performance on the VOT2015 dataset.