Multi-target object tracking method fusing object capture and recognition technology

Abstract

The invention provides a multi-target object tracking method and a multi-target object tracking device fused with object capture and recognition technology, which relate to the technical field of facerecognition and comprise the following steps: target capture: recognizing a to-be-tracked target of each frame, and taking a target position captured by a current frame recognition algorithm as an initial value of target tracking of a next frame; target comparison: comparing the similarity between the current frame target tracking result and the current frame target capturing result on the spatial position and the characteristic value by using an area overlapping method and an Euclidean distance; and target tracking: tracking the captured different targets at the same time through a multi-target tracking algorithm. Target tracking is assisted through a target recognition algorithm, multiple targets can be effectively tracked at the same time under the conditions of shielding and complex background interference, the target tracking accuracy is remarkably improved, and the problem of inaccurate tracking in the multi-frame continuous tracking process is effectively solved.

Description

technical field [0001] The invention relates to a multi-target object tracking technology, in particular to a multi-target object tracking method fused with object capture and recognition technology. Background technique [0002] In recent years, multi-object tracking has become a hot research topic in the field of computer vision. Currently existing tracking methods mainly include generative tracking methods and discriminative tracking methods. In the generative tracking method, the particle filter-based target tracking algorithm has shown great advantages under the influence of complex backgrounds, but there are still problems in the tracking process of multiple targets. If a target is blocked by other targets, it is very difficult to It is difficult to sort out the target relationship in such a situation so as to continue to track. Discriminative methods are generally better than generative methods, and can effectively distinguish background and foreground. [0003] At...

AI Technical Summary

Problems solved by technology

[0003] At present, many scholars have proposed discriminative tracking algorithms based on deep learning, such as tracking algorithms based on SAE (stack autoencoder); such as HCF (Hierarchical Convolutional Features) tracking algorithm, DeepSRDCF algorithm and C-COT algorithm. The deep convolution feature improves the discriminative correlation filter DCF (Disriminative Correlation Filter); the improvement of the kernel correlation filter (KCF) through the CNN feature is also a hot topic in the field of target tracking. Some related algorithms also combine support vector machines and Adaboost, etc. Traditional algorithms can achieve better results in tracking accuracy than traditional algorithms, and convolution features can also provide better results than manual features, but deep learning algorithms are not as simple as classification, identification, or detection methods. easy to succeed

Moreover, the current target tracking technology usually operates target detection and target tracking independently, but the continuous movement and changing angles of different targets in the video stream will reduce the accuracy of feature value comparison, and target comparison is prone to failure, resulting in tracking fail

[0004] Existing "a moving target recognition and tracking method and system based on a heterogeneous system on chip", the patent number is 201810980774.3, combining the background difference method and the frame difference method to obtain the target moving area, and track the target in the area through the Camshift algorithm. It is not easily affected by factors such as light changes and has good robustness, but the Camshift algorithm is easily disturbed by complex backgrounds, and it is difficult to deal with the situation where the background color or hue is close to the target

Images