Recently, on-device artificial intelligence (AI) solutions using mobile devices and embedded edge devices have emerged in various fields, such as computer vision, to address network traffic burdens, low-energy operations, and security problems. Although vision transformer deep learning models have outperformed conventional convolutional neural network (CNN) models in computer vision, they require more computations and parameters than CNN models. Thus, they are not directly applicable to embedded edge devices with limited hardware resources. Many researchers have proposed various model compression methods or lightweight architectures for vision transformers; however, there are only a few studies evaluating the effects of model compression techniques of vision transformers on performance. Regarding this problem, this paper presents a performance evaluation of vision transformers on embedded platforms. We investigated the behaviors of three vision transformers: DeiT, LeViT, and MobileViT. Each model performance was evaluated by accuracy and inference time on edge devices using the ImageNet dataset. We assessed the effects of the quantization method applied to the models on latency enhancement and accuracy degradation by profiling the proportion of response time occupied by major operations. In addition, we evaluated the performance of each model on GPU and EdgeTPU-based edge devices. In our experimental results, LeViT showed the best performance in CPU-based edge devices, and DeiT-small showed the highest performance improvement in GPU-based edge devices. In addition, only MobileViT models showed performance improvement on EdgeTPU. Summarizing the analysis results through profiling, the degree of performance improvement of each vision transformer model was highly dependent on the proportion of parts that could be optimized in the target edge device. In summary, to apply vision transformers to on-device AI solutions, either proper operation composition and optimizations specific to target edge devices must be considered.

Optimizing the performance of deep neural networks on embedded systems is a challenging task that requires efficient compilers and runtime systems. We propose a TVM-based approach that consists of three steps: quantization, auto-scheduling, and ahead-of-time compilation. Our approach reduces the computational complexity of models without significant loss of accuracy, and generates optimized code for various hardware platforms. We evaluate our approach on three representative CNNs using ImageNet Dataset on the NVIDIA Jetson AGX Xavier board and show that it outperforms baseline methods in terms of processing speed.

This paper presents a network-based automation system for the detection of metallic objects in food and livestock. A metal detector is a core equipment used for the inspection required by HACCP. Since the existing metal detectors are manufactured as a single-body equipment, it is difficult to take into account various user requirements for the system. In order to overcome the drawback, this study presents a network-based automation system for metal detector utilizing an industrial fieldbus and modular components. The proposed system can effectively consider the various customer requirements and control schemes. The proposed system can also achieve the improvement in speed and success rate of detecting metallic objects. The effectiveness of the proposed system is demonstrated through a various experiments.

In the machine vision inspection equipment, diameter measurement is important process in inspection of cylindrical object. However, machine vision inspection equipment requires complex algorithm processing such as camera distortion correction and perspective distortion correction, and the increase in processing time and cost required for precise diameter measurement. In this paper, we proposed the algorithm for diameter measurement of cylindrical object using the laser displacement sensor. In order to fit circle for given four input outer points, grid search algorithms using root-mean-square error and mean-absolute error are applied and compared. To solve the limitations of the grid search algorithm, we finally apply the singular-value-decomposition based circle fitting algorithm. In order to compare the performance of the algorithms, we generated the pseudo data of the outer points of the cylindrical object and applied each algorithm. As a result of the experiment, the grid search using root-mean-square error confirmed stable measurement results, but it was confirmed that real-time processing was difficult as the execution time was 10.8059 second. The execution time of mean-absolute error algorithm was greatly improved as 0.3639 second, but there was no weight according to the distance, so the result of algorithm is abnormal. On the other hand, the singular-value-decomposition method was not affected by the grid and could not only obtain precise detection results, but also confirmed a very good execution time of 0.6 millisecond.

Roundabouts have strengths in traffic flow and safety but can present difficulties for inexperienced drivers. Demand to acquire and analyze drone images has increased to enhance a traffic environment allowing drivers to deal with roundabouts easily. In this paper, we propose a roundabout traffic analysis system that detects, tracks, and analyzes vehicles using a deep learning-based object detection model (YOLOv7) in drone images. About 3600 images for object detection model learning and testing were extracted and labeled from 1 hour of drone video. Through training diverse conditions and evaluating the performance of object detection models, we achieved an average precision (AP) of up to 97.2%. In addition, we utilized SORT (Simple Online and Realtime Tracking) and OC-SORT (Observation-Centric SORT), a real-time object tracking algorithm, which resulted in an average MOTA (Multiple Object Tracking Accuracy) of up to 89.2%. By implementing a method for measuring roundabout entry speed, we achieved an accuracy of 94.5%.

This paper proposes a two-stage Water Shield Automation System (WSAS) to predict the possibility of river overflow and vehicle flooding due to sudden rainfall. The WSAS uses a two-stage Deep Neural Network (DNN) model. First, a river overflow prediction module is designed with LSTM to decide whether the river is flooded by predicting the river's water level rise. Second, a vehicle flooding prediction module predicts flooding of underground parking lots by detecting flooded tires with YOLOv5 from CCTV images. Finally, the WSAS automatically installs the water barrier whenever the river overflow and vehicle flooding events happen in the underground parking lots. The only constraint to implementing is that collecting training data for flooded vehicle tires is challenging. This paper exploits the Image C&S data augmentation technique to synthesize flooded tire images. Experimental results validate the superiority of WSAS by showing that the river overflow prediction module can reduce RMSE by three times compared with the previous method, and the vehicle flooding detection module can increase mAP by 20% compared with the naive detection method, respectively.