Vehicle Tracking Method Based on Machine Learning and Optical Flow

Abstract

The invention discloses a vehicle tracking method based on machine learning and optical flow. A vehicle model is obtained through one-off off-line training and used for detecting vehicle block mass Blobs in a video flow in real time, bidirectional pyramid optical flow tracking is performed on calculation characteristic point sets of all the vehicle mass block Blobs, and results of optical flow tracking in the forward direction and the backward direction are analyzed and filtered, so that multiple targets are stably and accurately tracked to form vehicle tracks. According to the complete vehicle tracking solution, the vehicle tracking method based on machine learning and optical flow can be widely applied to the fields of intelligent traffic, electronic polices, video monitoring, unmanned driving and others; by the utilization of the tracking method, a user can solve the classic problems in an existing tracking algorithm well, the multiple targets, such as long-period vehicle staying, size scale changing, shadowing, local shielding and touching, can be stably and accurately tracked; particularly, the vehicle tracking method has the good effects under the conditions of severe weather, a low illumination level and a high noisy point.

Description

technical field [0001] The invention relates to a vehicle tracking method based on machine learning and optical flow, which belongs to the vehicle tracking technology. Background technique [0002] Vehicle tracking has a very wide range of research and applications in the fields of intelligent transportation, video surveillance, and unmanned driving. Video-based vehicle tracking includes two modules: vehicle detection and tracking. At present, most vehicle detection methods use background difference-based methods, such as obtaining the background model through algorithms such as moving average, mixed Gaussian, codebook or Vibe, and then through difference. , binarization, morphological processing, and connected domain analysis to obtain the vehicle blob Blob. This method is based on pixel features, and it is difficult to solve problems such as sudden changes in light, long stays at red lights, adhesion, shadows, camera shake, etc.; the follow-up tracking algorithm can only ...

AI Technical Summary

Problems solved by technology

The traditional MeanShift tracking algorithm and particle tracking algorithm model the global features of the tracking area, and use a certain strategy to find the best candidate area. The disadvantage is that the feature is single (color histogram or LBP texture), and the target with monotonous color or multiple When the target sticks, it is easy to drift

The TLD tracking algorithm combines online learning and optical flow tracking, corrects tracking errors through online detection, and can track a single target for a long time, but this framework is not suitable for tracking multiple targets, especially multiple similar targets, and online learning time consuming

Images