دانلود کتاب کاربردهای پهپاد و حسگرها برای ناوبری و موقعیت‌یابی

In a wide range of AI (artificial intelligence)-enabled service fields such as human–computer interaction [1], traffic control [2], video surveillance [3], and augmented reality [4], object-tracking technology has drawn constant attention. Object tracking is largely divided into detection-free tracking and tracking-by-detection. Recent studies have used tracking-by-detection methodologies to realize MOT (multi-object tracking). However, this is a method of tracking classified objects in advance and the process of revealing the association between the detection results. Detection-free tracking, which allows users to track any object from the user point of view, can be a more useful technology for security and safety-related applications such as crime prevention and facility safety. VOT (visual object tracking) is a kind of detection-free tracking, which estimates the position of the user-defined target object in a series of video frames. In doing so, the estimated position in each frame is usually defined by the bounding box including the target object to be tracked. Without loss of generality, in order to secure the improved inference accuracy, DNN (deep neural network) models are getting bigger and more complicated [5]. Trackers with the latest technologies are also equipped with DNN models, and the computational complexity is also very high [6,7]. Therefore, tracking two or more user-specified objects in a video frame is even more computationally complex and makes the system very slow. The recent trend of VOT technology can be divided into Siamese-network-based and transformer-based studies. Siamese network structure tracks target object by computing the similarity between the target patch designated by the user and the search region of the video frames [6]. Transformer-based trackers conduct tracking by fusing the features of the target patch and search region of the frames using attention mechanism [7]. These two kinds of trackers have very different structures. Therefore, using the optimization technique of the same method is not meaningful to make up for the speed-performance deterioration incurred when tracking two or more objects. In order to maximize the execution speed of DNN, hardware accelerators specialized in specific DNN modules have been released [5]. However, these specialized accelerators do not guarantee their performance even in the new DNN structure. Therefore, edge devices and servers usually use GPUs to accelerate DNN modules. The DNN modules used in object trackers are also dependent on the GPU-specific libraries used by deep learning frameworks such as TensorFlow [8] and PyTorch [9], and the libraries are not optimized for various kinds of GPU hardware structures. Furthermore, these deep learning frameworks do not provide optimization techniques for two or more DNN modules to run parallel in the GPU even when the GPU is experiencing under utilization, which may lead the tracker performance to be suboptimal. To tackle the above-mentioned issues, we propose a software-based solution approach, which provides an efficient scheduling framework for the two well-performing object trackers running on edge devices and GPU-server computing systems. We first lay the groundwork for the proposed scheduling framework to optimally map workloads included in the tracker to computing units. To this end, an in-depth computational structure analysis is conducted on SiamRPN++, which epitomizes Siamese-network-based trackers, and CSWinTT, which best exemplifies transformer-based trackers. Particularly, we give most of our attention to the large-scale computational structure of MHA (multi-head attention), which transformer-based trackers have in common, from the DNN module perspective. Second, the proposed scheduling framework improves the tracking speed of two or more trackers when they are running together. This means that the tracking performance is improved when two or more objects are simultaneously tracked in a detection-free manner. The proposed scheduling framework is a system-level acceleration technology designed to be independent of the different structures of GPUs. Additionally, the approach proposed in this study can be applied to hardware accelerators other than GPUs. This is possible with only a library provided by the accelerator manufacturer.