Camera rotation and zoom-in and zoom-out identification method

Abstract

The invention discloses a camera rotation and zoom-in and zoom-out identification method, and relates to target detection based on video images, an intelligent event analysis system applied to video monitoring data in a traffic scene, and a traffic parameter calculation system. The method comprises the following steps: 1) extracting image data of each frame of a video, and carrying out feature point matching; 2) according to information accumulation of the matching feature points, fusing to generate a motion matching feature map; and 3) carrying out a deep learning multi-task enhanced segmentation classification algorithm on the motion matching feature map, and judging whether the camera has rotation and zoom-in and zoom-out actions according to a deep learning segmentation result. Compared with the prior art, the method adopts a general scheme, can realize camera rotation and zoom-in and zoom-out identification under multiple scenes, different weather conditions and different states, and has the characteristics of high identification precision and strong real-time performance.

Description

technical field [0001] The invention relates to target detection based on video images, an intelligent event analysis system applied to video surveillance data in traffic scenarios, and a traffic parameter calculation system, in particular, video data is converted to generate a single picture, and deep learning calculation is performed according to the characteristics of the picture , realizes the function of one model to complete multiple tasks, and completes the recognition calculation of camera rotation and zooming or zooming quickly, efficiently and accurately. Background technique [0002] The monitoring angle and range of the camera play an important role in practical applications, but in practical applications, some cameras may not be within the normal monitoring range, and because the manual cannot be found in time, the monitoring resources are wasted. [0003] The traditional camera rotation and zooming in and zooming identification method is based on the change of ...

AI Technical Summary

Problems solved by technology

[0008] 2. The target detection method based on deep learning has a better detection effect, but it takes time to detect in CPU mode and cannot achieve real-time results

[0009] 3. The efficiency is low, each frame needs to be calculated, and more resources are required

[0010] 4. The accuracy is low, the feature point extraction is uncertain, and the calculation amount of the feature point matching algorithm is relatively large

However, due to the complexity of the camera monitoring scene, the monitoring scene in different scenes and weather conditions is uncertain, and the scene between day and night is also quite different. It is difficult to unify many scenes with the traditional identification scheme of whether the camera is rotating and whether it is zoomed in or not. up, and the accuracy and efficiency are low

Images