Artificial intelligence (AI) cameras are edge devices with embedded graphics processing units that can run lightweight deep learning models. In traffic management applications, traffic flow and traffic incidents can be detected from roadside images with the use of AI cameras, and only detected high-level information is sent to the server to minimize the use of network bandwidth and server resources. However, because edge devices are computationally limited, models should be optimized before they are deployed to these AI cameras. In addition, environment-related parameters must be configured appropriately after model deployment. Thus, an AI camera management system is required. Consequently, in this study, we designed a deployment and management system for AI cameras; this system can perform model optimization and parameter configuration with ease. The main functions of this system involve 1) automatic modeling and code transfer, 2) the remote deployment of deep learning models, 3) the remote configuration of relevant applications, and 4) the presentation of analytical results on a graphical user interface. The performance of the developed system was investigated by using it to deploy traffic analysis models and visualize analysis results. The experimental results indicate that this system achieved all of its design goals.

Vehicle trajectory collection is critical for intelligent transportation systems and tasks such as driving behavior analysis, travel time measurement, and traffic planning. Object tracking through computer vision can be used to obtain vehicle trajectories; however, trajectory information collected from a single camera is limited because of the camera's limited field of view. In multi-target multi-camera (MTMC) tracking, multiple camera views are integrated to associate various trajectories to a single vehicle by matching the vehicle's appearance with the trajectories. These cameras might have overlapping or nonoverlapping fields of view. Trajectory information from MTMC tracking can be used for driving behavior analysis, traffic congestion estimation, and route planning. However, color tones and angles differ between cameras; thus, trajectory association is challenging. In MTMC tracking, appearance, spatial, and temporal information can be integrated to reduce identification failures. This paper proposes a trajectory association framework for travel time and traffic flow estimation. The effects of spatial and temporal information on performance were evaluated for the AI City Challenge dataset and a self-collected dataset. The F1 scores obtained for these two datasets were 0.917 and 0.897, respectively. The inclusion of spatial and temporal information improved the F1 scores by approximately 0.06-0.72. The errors for estimating travel time and vehicle behavior (turning or straight movement) were approximately 3 sand 15 %, respectively, for various camera angles.

AI cameras are edge devices that can execute lightweight deep learning models with embedded GPU devices. In traffic management applications, traffic flow and traffic incidents can be detected from roadside images by AI cameras, and only the detected high-level information needs to be sent back to the server to avoid network bandwidth consumption and spare server resources. However, due to limited hardware resources at edge devices, the models should be optimized for specific AI cameras before they are deployed. In addition, the environment-related parameters need to be configured properly after model deployment. These tasks call for an AI camera management system. In this research, we design a management and deployment traffic monitoring system which can accomplish model optimization and parameter configuration with ease. Except for the camera hardware installation, other main functions can be called remotely from the management system, including 1) Automatic modeling and code transfer generation; 2) Remote deep learning model deployment; 3) Remote application configuration; 4) Analysis result presentation with a graphical user interface. To validate our proposed system, the embedded GPU devices, including NVIDIA Jetson TX2 and AGX Xavier combined with roadside cameras, are used as the prototype of the AI cameras, and the deployment of intersection flow analysis models and the visualized analysis results are conducted by the proposed system. The experiments validate that the proposed management system achieves all the design goals.

Simultaneously controlling the signs of each intersection in the arterial road sections grouping based on the concept of arterial signal progression is the trend of traffic control over the past few years since traffic conditions are usually regional and diffuse, and the regulation of a single intersection sign can only achieve limited improvement. While adjusting a signal controller using traditional method, there are multiple parameters needed to be confirmed and set. To further control the intersection of other continuing sections, it is required to switch the signs sequentially by entering different intersection operation interfaces, which not only has a higher error rate, but also makes it difficult to relieve the traffic flow at the first moment since there will be a time lag in the change of the signal at each intersection. In order to solve this problem, we propose a sensing-active event script operation mode and a UI design based on arterial traffic signal control on the basis of the existing traffic log network. When the system detects changes in the performance of road sections (heavy traffic or congestion), it will trigger the number control adjustment mechanism, which links the website controlled by the real-time signal sent through the network system to the communication group in the operator's mobile phone. This paper takes Chiayi County as the field demonstration. Through our proposed sensing-active event script operation mode and UI-based arterial traffic signal control, the work efficiency of the operator has increased, and the rate of input error in the number of seconds of the sign has reduced that thus, relieved the traffic flow at once.

Traditional license plate recognition technology mostly uses traditional image processing methods to find out the characteristics of the license plate, and then crop and recognize the characters. The process needs to be modified due to the different environments, scenes and conditions. In recent years, many studies have implemented license plate and character recognition by using deep learning algorithms. Although it has a good recognition accuracy, the calculation speed still cannot reach the level of real-time recognition. This research proposes a real-time license plate recognition system based on YOLOv3, which uses deep learning model to realize the vehicle license plate recognition, lane identification and vehicle trajectory tracking. In this study, a web-based platform is established to present the result of license plate recognition and trajectory, and the streaming roadside video in the campus. In the platform, license plates of driving vehicles can be identified in real-time, and the user can search and track specific vehicle intuitively. In the experiment, the average accuracy of the system performs 84.3% in real-time license plate recognition, and 100% in lane identification. The system can process in 40 FPS, which can meet the level of real-time system. In the future, the system can cooperate with traffic access control in campus or community to improve the efficiency of traffic control.

Real-time traffic video streaming, such as roadside surveillance and aerial video, has been widely used in traffic monitoring nowadays. However, most of the traditional traffic data collection methods lack mobility that can only collect macroscopic data. In this paper, an intelligent traffic monitoring system based on an open source cooperative platform called SAGE2 was developed. Based on the integrated big screen TV wall of SAGE2, a map-based aerial traffic video streaming management interface was designed. In the image pre-processing section, it provides functions such as lens distortion removal, top view projection transforms, and video stabilization; simulate video streaming to provide instant and long-term micro-flow data collection. Micro-traffic flow data provides high-resolution information both in time and space which can be used to analyze the driving behavior of individuals and the public. Combined with the lane level map, it can provide a variety of visual vehicle flow presentations, such as intersection traffic distribution that can also be used to develop an innovative application in the future.

Heat shock protein 90 (HSP90) regulates the correct folding of nascent protein in tumor cells. Through the ATPase domain of HSP90, inhibition of its activity is a manipulation for anticancer treatment. Two series of anticancer compounds, flavonoids and YC-1 derivatives, were employed in this study. The reference ligand in the docking simulation showed the significant RMSD of 0.87 with respect to the template (PDB code: 1uy7). Six scoring functions (DockScore, PLP1, PLP2, LigScore1, LigScore2, and PMF) were employed to evaluate the binding affinity. The correlation coefficients (r2) between each scoring function and the biological activity were used to determine the accurate scoring function for virtual screening. The r2values were 0.878, 0.696, 0.395, 0.276, 0.050, and 0.187 for DockScore, LigScore1, LigScore2, PLP1, PLP2, and PMF, respectively. According to the accurate DockScore, most of flavonoids and YC-1 derivatives had the higher binding affinities to HSP90 than controls and built the important hydrogen bond with the key residue ASP93. The structure-based de novo design by using Ludi program was performed to increase the binding affinity. Final thirteen potential compounds had higher binding affinity than the original ones. These candidates might guide drug design for novel HSP90 inhibitors in the future.