Online multi-pedestrian tracking method and device based on Deep-Sort tracking framework

Abstract

The invention discloses an online multi-pedestrian tracking method and device based on a Deep-Sort tracking framework. The method comprises the steps: generating a target detection box through video frame input; calculating a cosine distance of the target detection frame; inputting the target detection box into the fused Vgg-16 network model, inputting the extracted characteristic value and track into the KCF, and calculating the Euclidean distance of the target detection box; and outputting a tracking result according to the cosine distance and the Euclidean distance in combination with the total distance of each target detection frame and the matching cascade mode of the Deep-Sort tracking framework. According to the method and device, Kalman filtering in a Deep-Sort tracking algorithm is changed into Gaussian kernel correlation filtering, so that a motion model can be established in a richer motion scene, meanwhile, the calculation time for estimating the pedestrian motion position is shortened, and the algorithm performance is improved; a peak sidelobe ratio is provided by aiming at a drifting phenomenon generated by tracking pedestrian movement by Gaussian kernel correlation filters, and a plurality of correlation filters are connected to achieve higher tracking accuracy and precision.

Description

technical field [0001] The invention belongs to the technical field of multi-pedestrian tracking, and relates to an online multi-pedestrian tracking method and device based on a Deep-Sort tracking framework. Background technique [0002] Multi-pedestrian tracking is one of the common scenarios in the field of video surveillance, which means that pedestrian motion video is input into the multi-pedestrian tracking algorithm to obtain the trajectory of each pedestrian, including the start and end of the trajectory. These pedestrian trajectories can be further analyzed according to practical purposes, such as the analysis of abnormal behavior of pedestrians. Therefore, multi-pedestrian tracking plays a linking role in surveillance video analysis, making surveillance information more valuable. With the vigorous development of deep learning, a large number of multi-pedestrian tracking methods based on deep learning have also emerged rapidly, such as C-COT, DeepSort, MOTDT, DeepMOT...

AI Technical Summary

Problems solved by technology

[0008] (1) Deep-Sort uses the Kalman filter to establish the pedestrian motion model. This method is simple in principle and easy to calculate, but it has certain requirements for the motion of pedestrians: the motion of pedestrians must be in a linear system. Pedestrian tracking will be lost when occluded

[0009] (2) Because the target frame of KCF has been set during the tracking process, the size has not changed from the beginning to the end, but in the general tracking sequence, the target size is difficult to keep constant for a long time, which will cause the tracker to The target frame drifts during the tracking process, which is the so-called tracking drift

However, when tracking in a real scene, although the appearance may remain stable in the short term, in the long run, interruptions in the external environment may occur from time to time, and sufficient attention must be paid to these changes, such as occlusion, lighting and other issues

Images