Multi-view Indoor Pedestrian Tracking Method Based on Motion Behavior Pattern

Abstract

The invention discloses a multi-view indoor pedestrian tracking method based on a movement behavior mode. The method comprises the steps of predicating the target movement state through a state space model and observing the predicated movement state through an observing model to obtain the tracking result, wherein the predicating includes that calculating the resultant force applied to a target in a testing video frame at the time t(i) ( / i), calculating the probable force of the target in the testing video frame at the time (i) ( / i), and predicating the target movement state at the next time according to the probable force of the target. With the adoption of the method, the validity of the pedestrian tracking information can be improved, and the robustness and accuracy of pedestrian tracking can be improved.

Description

technical field [0001] The invention belongs to the technical field of image processing, in particular to a pedestrian tracking method, and more specifically to a multi-view indoor pedestrian tracking method based on motion behavior patterns. Background technique [0002] Visual information is the main channel for the human body to obtain external information, of which motion information is an important part of it, and a large amount of important and meaningful visual information is contained in motion. Based on the important value of moving target analysis in practical applications, and the target analysis system established on the basis of analysis, its performance depends on the extraction and analysis of target motion information. Therefore, for moving target tracking technology research is of great significance. [0003] The target tracking is to obtain the position of the target by analyzing the collected data about the target movement, and then obtain the trajectory ...

AI Technical Summary

Problems solved by technology

[0011] 1) The tracking method under the particle filter framework mainly uses the Monte Carlo method to solve the integral operation in the Bayesian estimation based on the theorem of large numbers, which can better solve the target tracking problem in nonlinear and non-Gaussian distribution systems, but The above method randomly generates candidate particles, only guarantees the diversity of particle sampling, and does not select a better reference distribution to improve sampling efficiency and increase the amount of calculation, and the randomly generated particles do not consider the specific motion behavior of the target, which has a certain Blindness; the pedestrian tracking method based on manifold learning and sparse representation needs to manually select the pedestrian target and reduce the video frame to a uniform size before tracking, which limits the application of the tracking method

[0012] 2) The method of segmenting and combining the clumps of the foreground image is susceptible to interference from factors such as illumination changes and occlusion of the human body, which makes the segmentation of the human body area incomplete and prone to tracking drift. For counting, the association of target pedestrians between different frames cannot be achieved, and continuous tracking cannot be achieved

[0013] 3) In the above-mentioned tracking method, the visual tracking is transformed into the target matching problem in continuous video frames. Its essence can be regarded as a local matching optimization problem in a continuous local space, which can realize the detection and tracking of the target, but it needs Obtaining images of all scales for recognition increases the amount of calculation. At the same time, by calculating the similarity of targets between adjacent frames to match and associate different target pedestrians, it is difficult to obtain accurate matching in scenes with relatively many pedestrians. Appearance and disappearance cases are also not handled efficiently

[0014] 4) The limitations of pedestrian tracking methods based on improved random forest and HOG-LBP description are as follows: all such methods need to establish target classifiers, but there are three main problems. First, the construction of classifiers requires a large number of positive and negative samples. Learning, how to select samples is a key issue; second, the computational complexity is high, the amount of calculation is large, and it is difficult to meet real-time needs; If the range is too small, the target tracking accuracy will be affected, and the search efficiency will not be reduced because the range is too large. Further theoretical research is still needed.

Images