Infrared video multi-target tracking method

Abstract

The invention discloses an infrared video multi-target tracking method which mainly solves the problems that in the prior art, during target detection, the detection precision is low, and a model is prone to over-fitting. The method comprises the following steps: selecting people and vehicle targets in an infrared image, carrying out histogram equalization on the infrared image, and constructing a training data set; modifying a backbone network on the basis of an existing RFBNet network, and constructing a target detection network; training a target detection network by using the training data set and adopting a gradient descent method; inputting a to-be-tracked infrared video into the trained target detection network to obtain a target detection result in the video; and selecting a DeepSORT multi-target tracking algorithm to carry out data association on the target detection result, establishing human and vehicle movement tracks, and obtaining an infrared multi-target tracking result. According to the method, the over-fitting degree of the target detection model is reduced, the tracking precision is improved, and the method can be used for multi-target real-time tracking of pedestrians and vehicles in a complex infrared scene.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to an infrared video multi-target tracking method, which can be used for real-time tracking of multiple targets of pedestrians and vehicles in complex infrared scenes. [0002] technical background [0003] In recent years, deep learning methods have been mainly used in target recognition and tracking tasks. This method has powerful feature modeling capabilities, which mainly fall into the following two categories: [0004] The first category is to combine convolutional features with correlation filters. For example, Danelljan et al. proposed the C-COT algorithm. By learning in a continuous resolution sequence and creating a time-domain continuous correlation filter, feature maps of different resolutions can be used as input to the filter, making traditional features and depth Features can be deeply combined. The disadvantage of this type of method is that the tracking sp...

AI Technical Summary

Problems solved by technology

Because infrared video is different from the color video used in general deep learning, its target is more easily affected by the surrounding environment, and the appearance of the target often has a large range of changes, mainly manifested in changes in outline and gray distribution, making tracking difficult

At the same time, due to the low resolution of infrared video and the lack of available channel information, the use of a neural network with a large number of parameters and too deep depth can easily lead to over-fitting of the model and reduce the tracking effect

Images