Video privacy data fuzzification method running on edge device

Abstract

The invention discloses a method for fuzzifying video privacy data running on edge equipment. The method comprises the steps of algorithm model design, algorithm model building, model optimization, model quantification and acceleration, and model migration to mobile terminal equipment for running. The privacy automatic fuzzy system has the beneficial effects that the privacy automatic fuzzy system which is low in computing resource requirement, high in computing speed, capable of running on the mobile terminal and independent of other server resources is provided, and the privacy of other people in the public video is protected; a multi-target tracking algorithm is adopted to track an object, so that a fuzzy object can be specified or a motion track of a specified target can be recorded; the algorithm model is quantized, optimized and accelerated by using TensorRT, the deployment difficulty is reduced, and meanwhile, when the algorithm runs, three threads are adopted for processing the three parts of preprocessing, model reasoning and post-processing respectively, so that the running efficiency is improved.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a method for blurring video privacy data running on an edge device. Background technique [0002] The scope of application of monitoring and other camera equipment is more and more extensive, such as the monitoring equipment of the community, or the kitchen environment is disclosed through the camera, or the video needs to be collected through the camera and uploaded to the Internet; at the same time, people are paying more and more attention to privacy protection. Although the disclosure of surveillance can help people supervise the kitchen sanitation environment, it also leads to the disclosure of chefs' privacy; posting videos on the Internet is an efficient way of dissemination, and at the same time, the efficient dissemination and searchability of network data are also important. It is easier to cause privacy leakage of others in the video. [0003] The ...

AI Technical Summary

Problems solved by technology

[0002] The scope of application of monitoring and other camera equipment is more and more extensive, such as the monitoring equipment of the community, or the kitchen environment is disclosed through the camera, or the video needs to be collected through the camera and uploaded to the Internet; at the same time, people are paying more and more attention to privacy protection. Although the disclosure of surveillance can help people supervise the kitchen sanitation environment, it also leads to the disclosure of chefs' privacy; posting videos on the Internet is an efficient way of dissemination, and at the same time, the efficient dissemination and searchability of network data are also important. It is easier to cause privacy leakage of others in the video

[0003] The existing processing methods are mainly late-stage desensitization processing, and post-stage desensitization requires manpower processing, and post-processing is not suitable for online monitoring equipment, and most of the processing methods are frame-by-frame processing, with low efficiency

[0004] In computer vision technology, segmentation technology can quickly obtain the mask of all people in the video. Segmentation technology can be divided into semantic segmentation technology and instance segmentation technology; but the difference between semantic segmentation and instance segmentation is that semantic segmentation refers to pixel-level Classification, classifying each pixel in the image, taking people as an example, that is, judging whether each pixel is a part of a person; instance segmentation is a further distinction based on semantic segmentation, not only distinguishing between people and backgrounds, but also distinguishing People and people; therefore, semantic segmentation can achieve fast segmentation, but it is not flexible enough, and when the target object accounts for a small proportion of the image, its segmentation accuracy is low, and the instance segmentation technology is more flexible, but its calculation load is also larger. More processing is required to run on low-computing devices; at the same time, instance-level processing can be easily added to target tracking technology

[0005] Convolutional neural networks can achieve high-precision instance segmentation, but instance segmentation has a huge amount of computation, and most of them can only be run on the server side, and it is difficult to directly deploy to mobile devices; and with the rapid development of society, the The number is also growing explosively; it is difficult to rely on the server alone to handle large-scale instance segmentation algorithms. At the same time, there is a risk of privacy leakage in the process of uploading video data by using the server. Therefore, an edge device is required. The method of edge computing is used to process these data in real time. Edge computing refers to the side close to the source of objects or data, and its applications run on edge devices to generate faster service responses and meet real-time business requirements.

Images