Bounding Box Aware Edge-Cloud Collaborative Method for Multiple Object Detection

The demand for real-time video processing from edge devices including surveillance cameras and smartphones has been increasing. While edge processing power is improving with lighter recognition models and smaller GPUs, achieving high-performance recognition remains a challenge due to limited computational resources. To address the issue, collaborative recognition systems between the edge side and the cloud side are crucial. Previous approaches such as the Edge-Cloud Net (ECNet) have been proposed but they faced challenges in optimizing data transmission because of the large data size of frame images. In this paper, we propose a novel edge-cloud collaborative method for video multiple object detection. This system integrates an original lightweight edge side model that combines YOLOv3 and YOLOv3-tiny and compresses intermediate features before the transmission. Our approach improves the trade-off between transmission amount and detection accuracy, particularly at low bit rates. This approach also focuses on offload controlling based on detected bounding boxes from the edge side model and it enhances the trade-off compared to the previous method.