Infrared Imaging Target Recognition And Tracking Based On Learning Algorithm

Target identification and tracking have been the subject of many excellent research results in the field of pattern recognition,but the recognition rate,speed of operation,and stability of IR imaging targets have yet to be improved.In this paper,we start from the application of infrared imaging in road surveillance and use machine learning-based algorithms for target identification and tracking.First,this paper analyzes the principles and characteristics of infrared imaging and improves the method for detecting the significance of infrared targets based on reconstruction optimization.First,the a priori knowledge of IR imaging is analyzed,the appearance-based IR significance enhancement is introduced to improve the contrast between target and background,then the multi-scale contrast enhancement algorithm is used to improve the target significance,and due to the difference in brightness between target and background,a hybrid Gaussian model is used to construct a gradient probability distribution map of IR images to separate target and background;finally,the energy equation is constructed based on the a priori enhancement of IR imaging and target significance,and the optimal solution is calculated using the random wandering algorithm.Experiments on a large number of traffic road monitoring images have shown that the improved significance detection algorithm outperforms other algorithms in terms of accuracy,recall and average error.Second,this paper improves the way the Support Vector Machine(SVM)classifier is evaluated for cost.For traditional SVM classifiers,the cost of sample misclassification is the same,training the classifier to evaluate the classifier by looking for hyperplane segmentation functions that minimize the total misclassification cost and by error rates.However,due to the complexity of the samples themselves,it is not unreasonable to use the same cost assessment cost for different samples,and this paper adopts minimal empirical risk and structural risk evaluation criteria to optimize the classifier in terms of both the total cost of the training set and model complexity to reduce the high cost error of misclassified samples.Improved SVM classifier experiments on IR imaging datasets show that the method reduces the number of classifier misclassification samples and is generalizable across datasets.Finally,the video frame image maximum response position is calculated by correlation filter to locate the target and target tracking is achieved based on the online learning tracking algorithm Kernel Correlation Filter(KCF).Particle filter(PF)is used to build a target motion model,which solves the tracking problem when the nuclear correlation filter cannot handle target masking,and provides the target detection region of the nuclear correlation filter to realize the secondary tracking problem after target masking.Experimental validation in the open OTCBVS dataset and real road scenarios shows that the algorithm can correctly identify road vehicles and achieve stable real-time tracking of vehicles in IR surveillance video.The correct target identification rate using the SVM classifier reached 93.27%,the tracer ran at 14.56 FPS,and the average distance accuracy and average overlap accuracy of the tracer were higher than other algorithms.