To address the trade-off between accuracy and latency in real-time video analysis tasks and further enhance the system,a cloud-edge collaborative video stream analysis framework based on saliency detection was proposed.The saliency of different regions was accurately measured by calculating the gradient of the inference result on the video frame,and the perception area of the macroblock(the basic unit of video coding)was constructed by combining the context features of the surveillance camera.By establishing these perception areas,the edge end could perform different levels of compression and filtering on the content of dif-ferent areas in the video frame,thereby reducing bandwidth consumption during transmission.Experimental results demonstrate that the framework can significantly reduce bandwidth consumption and latency with only slight accuracy degradation.
维普
万方数据