Edge cloud collaborative deep learning target detection method based on target tracking acceleration

Abstract

According to the edge cloud collaborative deep learning target detection method based on target tracking acceleration provided by the invention, the problem that the real-time performance of the target detection problem cannot be guaranteed is solved. The method comprises the following three stages of processing: in the first stage, a key frame selection method is provided by using a self-adaptive key frame algorithm on an edge node, and key frames are selected in the same deep learning model only by extremely low computing resource cost; and a second stage, at the cloud, high-precision target detection is peformed by using the edge screened data and a high-precision classification model; and a third stage, at an edge end, rapid tracking is carried out through classification of key frames and a frame marking result by utilizing a twin network, and according to the method, data screening aiming at video target detection is realized by utilizing an adaptive key frame algorithm, and meanwhile, compromise between model precision and time delay consumption is realized. A reliable scheme is provided for solving the video target detection problem of the edge cloud collaborative deep learning model.

Description

technical field [0001] The present invention belongs to the field of artificial intelligence, and specifically relates to a fast online video object detection method that utilizes in a collaborative manner accurate object detectors on the cloud and lightweight object trackers on devices with limited resources at the edge of the system. Background technique [0002] Video object detection plays an important role in a growing number of smart city applications that require intelligent video analytics. However, the dominant methods either suffer from the long response time of end-to-end video object detection or their offline nature, making them unsuitable for delay-sensitive analysis of video streams. Traditional centralized cloud computing is often used to train high-precision deep learning models, such as deep neural networks. However, the delay of data uploading to the cloud makes the detection speed on the cloud slow. Using the distributed edge computing paradigm, the edge...

AI Technical Summary

Problems solved by technology

However, the dominant methods either suffer from long response times for end-to-end video object detection or suffer from their offline nature, making them unsuitable for latency-sensitive analysis of video streams

Traditional centralized cloud computing is often used to train high-precision deep learning models, such as deep neural networks. However, the delay of data uploading to the cloud makes the detection speed on the cloud slow

Images