Low-identification Target Recognition Algorithm Based On Feature Fusion

With the development of modern machine learning and deep learning algorithms,the accuracy of environmental perception technology for pedestrian,vehicle and other target recognition in an ideal environment has been continuously improved.However,the environment detection technology has a low detection accuracy rate for many special scenes in real life,such as rainy and low-identification environments at night.Therefore,breaking through the accuracy level of low-recognition target detection is the key to the practical application of environmental perception technology.The characteristics of low-identification targets include: low contrast between the target and the surrounding environment,no obvious and complete contour features of the target,or a target pixel accounting for less than 30% of the entire image pixels.This paper focuses on the recognition algorithm of pedestrian and vehicle targets in the dim city street light scene.Using the idea of multi-feature fusion,on the basis of conventional single-mode color camera detection targets,by increasing the target feature dimension to improve detection and recognition Accuracy.The main work of this article is as follows:(1)Research on low-identification target detection and recognition algorithms that incorporate multi-dimensional perceptual features.Aiming at the problem of unclear,inconspicuous,and incomplete target features in a low-recognition environment with dim light,a traditional machine vision target detection algorithm that is different from a single threshold is proposed.The algorithm is improved from two aspects: detection and recognition.Firstly,a mean-shift clustering algorithm based on YCb Cr color space is proposed at the detection stage to enhance the region of interest of the target.Then in the target recognition phase,the target features extracted by the HOG algorithm and the 4-layer CNN neural network are fused to increase the accuracy of target classification by subsequent SVM.(2)Research on The dual YOLO v3 recognition low-identification target algorithm based on color and infrared sensors.The accuracy of the above-mentioned machine vision algorithm is low,and the accuracy of low-recognition target recognition cannot be improved by merely improving the YOLO v3 algorithm.In order to solve the above problems,this paper proposes a dual YOLO v3 algorithm.In this algorithm,color images and thermal infrared images are simultaneously input to the improved network for feature extraction,and the features of the two modalities are fused at a certain layer of the network.Four fusion schemes are designed in this paper to get the optimal network fusion model.In order to further improve the accuracy of the algorithm for low-recognition target recognition,this paper also introduces Focal-loss and GIOU to further optimize the parameters of the classification layer.(3)Study the step-by-step training method and parameter setting of the dual-modal YOLO v3 algorithm.Due to the low number of image pairs and thermal images in modern open source datasets,therefore,this article has built a synchronous acquisition system of dual modal data sets to collect multiple types of low-identification target data sets in different scenarios.Moreover,the data set is processed for registration,enhancement,and labeling for subsequent training of the dual-modal neural network.This paper also proposes a step-by-step network training method,which can get better results than traditional training methods.The experimental results show that the recognition accuracy of low-identification targets incorporating multi-dimensional perception features is significantly improved,but targets that are distant and blocked by halo in the surrounding environment cannot be identified;The dual YOLO v3 neural network algorithm is used to simultaneously identify The low-identification targets collected by the infrared and color sensors can well recognize the occluded long-distance targets,and the accuracy is significantly improved;Using step-by-step training methods for network training,the network algorithm parameters are adjusted to be optimal.