A Deep Learning-Based Multi-Pedestrian Tracking Method

Abstract

The invention provides an online multi-pedestrian tracking method based on deep learning, which mainly involves tracking multi-pedestrians in complex scenes by means of deep learning. The method includes: proposing a multi-layer feature fusion network to extract the appearance features of the target. Use a high-precision object detector to detect the position of the object in the video, and extract the appearance features of the object. For different state trajectories, considering the confidence of appearance features and motion features, the difference metric matrix is ​​calculated in different ways. According to the confidence of different state trajectories, combined with the metric matrix, the Hungarian algorithm is used to realize the hierarchical data association of detection and trajectories. Finally, the trajectory information is updated according to the association result. The invention fully considers the confidence of the target motion feature and the appearance feature, and reasonably utilizes the two features to calculate the difference metric matrix. At the same time, the data association method of hierarchical association can better deal with the problems of pedestrian false detection and occlusion, and effectively reduce the number of trajectory ID transformations.

Description

technical field [0001] The invention relates to the multi-target tracking problem in the field of deep learning, in particular to a multi-pedestrian tracking method based on deep learning. Background technique [0002] Video-based tracking algorithms have always been an important branch of computer vision research, and emerging new technologies such as correlation filtering, convolutional neural networks, etc. have also been rapidly integrated into the field and significantly improved the average level of the field. With the emergence of a large number of deep learning technologies, the field of computer vision has once again achieved rapid development. The tracking algorithms based on deep learning are constantly improving and have a wide range of application scenarios. For example: intelligent video monitoring and control, abnormal behavior identification and analysis, human-computer interaction, medical image analysis, pedestrian flow analysis in public places, etc. [0...

AI Technical Summary

Problems solved by technology

Traditional methods are prone to trajectory drift problems in complex scenes with many targets, and the ability to deal with occlusion problems is also limited.

Images