A Research On Infrared Weak Small Object Correlation Filter Tracking Algorithm Based On Deep Learning

The object tracking method using infrared thermal imaging technology is widely used in precision guidance,intelligent monitoring,and automatic driving due to its strong anti-interference,adaptability and all-weather working characteristics.However,the low contrast and local detail blur inherent in infrared images can cause the feature information of the tracking object to be weakened.The traditional correlated filtering infrared object tracking algorithm is limited by the weak representation ability of the shallow features of the object,boundary effects and deformation occlusion,etc.,and the object robust tracking cannot be achieved.The object tracking algorithm based on the deep convolutional neural network can effectively solve the robustness of tracking because its extracted deep features contain more semantic information.Therefore,this paper will study the infrared image contrast optimization and weak detail enhancement,the antiocclusion research of object tracking and the application of depth features.The main innovations and research contents of this paper are as follows:(1)Aiming at the problem of texture blur and low contrast of infrared weak objects,a weak detail enhancement method for infrared images is proposed.On the basis of the guided filtering layering,the improved fast multi-scale median filtering is applied to the detail layer to suppress the detail layer noise;the background layer fusion global histogram information is used to improve the CLAHE enhancement effect;finally,the multi-scale de-artifact method is used to optimize Fusion image.Experimental results show that the proposed enhancement algorithm effectively improves the local contrast and detail definition of infrared images.While ensuring real-time,the information entropy and PSNR indicators have been greatly improved.(2)Aiming at the problem that the infrared object is occluded and interfered by the background clutter,an improved infrared object tracking algorithm based on the idea of fusion re-detection is proposed.For the occlusion problem,the multi-peak energy detection of the object response is established,and the learning rate and the high confidence update of the model are implemented based on the detection result to solve the model drift caused by the occlusion and to punish the relevant response value of the background region.Resolve interference from similar objects.Combined with the SVM classifier to achieve the recapture after the object is lost.The experimental results show that the proposed filter tracking algorithm is stable and robust,and can achieve object recompensation when the object is lost.Compared with the LCT algorithm,the accuracy and success rate are improved by 5.6% and 4.1%.(3)Aiming at the robust deep semantic features and spatially accurate shallow texture features extracted by CNN,a stable correlated filtering object tracking algorithm based on multi-layer depth features is proposed.The algorithm adaptively fuses the multilayer features to enhance the robustness of the tracking;combined with the time domain context distance constraint and high confidence update to achieve the object re-detection.Use the Kalman filter and confidence results to solve the object loss problem.The experimental results show that the proposed object tracking method based on depth features can achieve robust tracking of small objects.Compared with other deep learning algorithms,the accuracy and success rate have been improved to some extent.