Edge-assisted object detection on mobile devices via the reuse of bounding boxes

Abstract

Object detection is a type of technology that locates and classifies meaningful objects such as people and cars in digital images or videos. Recently, the field of object detection technology made a great breakthrough in conjunction with the significant development of Deep Learning (DL). However, it is difficult to realize computation-intensive object detection on mobile devices equipped with low-spec hardware due to the long latency. To solve this problem, previous studies proposed solutions that offload the object detection task to a more powerful edge server. However, these studies did not attempt to exploit earlier object detection results. Therefore, this thesis presents Edge-assisted Object detection on Mobile devices via the Reuse of bounding boxes (EOMR), a system that enables efficient and low-latency object detection on mobile devices using previous detection results. EOMR measures frame similarity levels based on frame differences only for regions of previous detection results. Then, EOMR performs only one of the three tasks: detection result reusing, object tracking, and offloading for the current frame based on the calculated frame similarity. We evaluated our system with three videos. Experiment results show that EOMR significantly reduces the average processing latency per frame while tolerably reducing the detection accuracy compared to the baseline, which only performs offloading.