Scheduling Framework for Accelerating Multiple Detection-Free Object Trackers

In detection-free tracking, after users freely designate the location of the object to be tracked in the first frame of the video sequence, the location of the object is continuously found in the following video frame sequence. Recently, technologies using a Siamese network and transformer based on DNN modules have been evaluated as very excellent in terms of tracking accuracy. The high computational complexity due to the usage of the DNN module is not a preferred feature in terms of execution speed, and when tracking two or more objects, a bottleneck effect occurs in the DNN accelerator such as the GPU, which inevitably results in a larger delay. To address this problem, we propose a tracker scheduling framework. First, the computation structures of representative trackers are analyzed, and the scheduling unit suitable for the execution characteristics of each tracker is derived. Based on this analysis, the decomposed workloads of trackers are multi-threaded under the control of the scheduling framework. CPU-side multi-threading leads the GPU to a work-conserving state while enabling parallel processing as much as possible even within a single GPU depending on the resource availability of the internal hardware. The proposed framework is a general-purpose system-level software solution that can be applied not only to GPUs but also to other hardware accelerators. As a result of confirmation through various experiments, when tracking two objects, the execution speed was improved by up to 55% while maintaining almost the same accuracy as the existing method.