Research On Infrared Video Object Tracking Algorithm Based On Deep Learning

With the rapid development of artificial intelligence technology,image algorithms based on deep learning are widely used in various of areas.Among them,object tracking is a blend of various theory s and algorithms of the various fields,such as image processing,machine learning and optimization.It’s also the premise and foundation of achieving higher levels of image understanding,such as target behavior recognition tasks.Infrared object tracking algorithm can break the limitation of object tracking scene depending on visible light,and can also track objects in dark night,rain,snow,wind and frost and other harsh environments,so as to realize the application of video surveillance at night,assisted driving at night,maritime rescue and other scenes.Due to the inherent process of infrared imaging system,drastic change of brightness may occur in infrared video sequences under some scenes,leading to relatively drastic feature changes between adjacent frames,which is not conducive to the discrimination of target features by target tracking algorithm.Based on the principle of infrared image imaging,this paper explored the phenomenon of brightness change in infrared imaging,and proposed a the correction method to smooth the brightness change and improve the tracking performance.In addition,a visual attention mechanism is introduced to extract more discriminant features from the convolutional neural network used for feature extraction module in target tracking algorithm.The main works of this paper are as follows:1.The current object tracking model is usually trained based on RGB images.In this paper,by means of transfer learning,the features of RGB images are used as pretraining and the features of infrared targets are integrated into the feature extraction module by using the sequence fine-tuning of infrared image training set,so as to improve the tracking effect of the model on infrared video sequences.2.Aiming at the phenomenon of background brightness drastic change in infrared video sequences when high-temperature objects enter or leave the scene,the classical method of Kalman filter is introduced.Based on the principle of infrared imaging,the background brightness is modeled using Kalman filter to correct the brghtness,which not only improves the visual effect but also improves the tracking performance.3.On account of the feature extraction module of convolutional neural network,the visual attention mechanism,including channel attention and spatial attention,is introduced to make the network locate the information of interest,suppress useless information and optimize the feature extraction process,thus improving the tracking effect.