論文

One major research problem regarding mobile sensing systems is how to accelerate the processing speed which is bottle-necked by the deep-learning-based object detection. The focus of this paper is directed towards a typical mobile sensing scenario wherein sequences of frames containing interested objects are sparsely dispersed throughout the video stream. In light of this, we propose a stochastic scheduling algorithm named JumpQ. In the case of consecutive negative detections, JumpQ reduces the probability of detection, while in the case of positive detections, JumpQ promptly returns to frame-by-frame detection and retraces the buffered frames to detect the objects. Our experiment reports that the JumpQ algorithm accelerates processing speed by over 100%, all while incurring a negligible impact on sensing accuracy.

Edge computing plays a key role in smart-city deployments. Deploying AI models on edge devices reduces latency and privacy concerns. This work proposes QPilot, a reinforcement-learning-based adaptive model-switching method that dynamically selects detection models of different sizes for multi-object tracking (MOT) inference to balance efficiency and accuracy according to device compute capacity and scene complexity. In object-sparse scenarios, it improves tracking accuracy with only a small additional computational cost and achieves an optimal efficiency-accuracy trade-off in MOT. We validated the method on multiple hardware platforms, showing strong cross-device adaptability as well as feasibility and promise for edge environments.

We propose a 3-dimensional(3D) CO2 concentration analyzing system to monitor the ventilation of the closed space. this system employs a 3D scan to reduce the labor associated with creating a 3D space model. This research uses multiple CO2 sensors and considers the model to map point CO2 concentration data to the spatial CO2 distribution. We experimented with evaluating the accuracy of the model in the actual bus space.

Crowdedness of buses is playing an increasingly important role in the disease control of COVID-19. The lack of a practical approach to sensing the crowdedness of buses is a major problem. This paper proposes a bus crowdedness sensing system which exploits deep learningbased object detection to count the numbers of passengers getting on and off a bus and thus estimate the crowdedness of buses in real time. In our prototype system, we combine YOLOv5s object detection model with Kalman Filter object tracking algorithm to implement a sensing algorithm running on a Jetson nano-based vehicular device mounted on a bus. By using the driving recorder video data taken from real bus, we experimentally evaluate the performance of the proposed sensing system to verify that our proposed system system improves counting accuracy and achieves real-time processing at the Jetson Nano platform.

The combination of object detection, tracking, and counting has become a widely used method. The position and angle of cameras can vary according to the deployment scenarios, which affects counting accuracy. Traditional approaches often rely on manually pre-defined counting lines or regions-of-interest (ROIs), which are static, environment-specific, and difficult to generalize. To overcome these limitations, we propose AdaLine, an adaptive, perspective-aware algorithm that learns the optimal counting line from object trajectories, thereby enabling adaptation to diverse environments and camera viewpoints without manually defined counting lines. AdaLine adapts automatically as the scene evolves by clustering incoming trajectories with K-means, selecting the most stable line candidate, and smoothing it with an exponential moving average. Experimental evaluations across different scenarios and camera settings show that AdaLine achieves better performance in terms of accuracy, stability, and applicability. Our approach offers a scalable, real-time configuration-free solution for object-detection-tracking systems.

Combining object detection, tracking, and counting has become a widely-used method. The position and angle of cameras can vary according to the deployment scenarios, both of which affect object counting. In current practice, we have to set the position of the counting line manually. However, determining the optimal position for each scenario remains a challenge. In this paper, we propose the AdaLine, an adaptive optimization algorithm for object counting lines in detection-tracking-counting sensing systems. AdaLine calculates the movement characteristics of objects to automatically generate a counting line suitable for deployment scenarios. By using the driving recorder video data taken from real buses, we experimentally evaluate the sensing performance of the proposed algorithm and verify that our proposed algorithm improves counting accuracy.

Crowdedness sensing of buses is playing an important role in the disease control of COVID-19 and bus resource scheduling. This research analyzes the relationship between carbon dioxide concentration, bus environment and the number of passengers by linear regression. Our prototype system collects the data of bus environment and carbon dioxide concentration to estimate the number of passengers in real time. By collecting the sensing data from a shuttle bus of university campus, we experimentally evaluate the feasibility and sensing performance of the crowdedness estimation model.

Forecasting household waste generation using conventional methods can present challenges due to the substantial variability and uncertainty in the process. Furthermore, previous studies focused on forecasting household waste generation at municipal or national levels may not be directly applicable to on-site waste collection processes. The objective of this research is to attain daily-level predictions of household waste generation and assess the advantages of using a leading-edge deep learning approach over conventional methods. We applied Multi-variable Long Short-Term Memory (LSTM) neural network utilizing a granular garbage disposal database for forecasting. This database is curated from a garbage disposal sensing platform currently operational in three cities within the Kanagawa prefecture, Japan, with plan for operation until 2025. Additionally, relevant web applications based on findings from this research will be developed for data visualization and routine optimization.

Deep learning-based object detection has seen a surge in applications for sensing systems on mobile devices. In this context, objects are identified and tracked across video frames, facilitating the calculation of associated events of interest. A significant research challenge refers to the acceleration of processing speed, which is constrained by deep learning-based object detection due to its intensive resource requirements. This paper focuses on a typical mobile sensing scenario, wherein sequences of frames containing objects of interest are sparsely dispersed throughout the video stream. Given that many of the frames lack objects, allocating substantial computational resources to detect them becomes inefficient. In light of this, we propose a stochastic scheduling algorithm, JumpQ. JumpQ performs per-frame detection when anticipating the presence of objects in the current frames. Consecutive negative detections prompt a transition to intermittent detection with a probability that undergoes further decay if the negative detection persists until reaching a predefined limit. Upon a positive detection, JumpQ swiftly reverts to per-frame detection and retraces a specific number of previously buffered frames to ensure the inclusion of potentially missed true frames. A comprehensive experimental study using the garbage bag counting technique was conducted to show the efficiency of JumpQ in accelerating the processing speed by nearly 1.92 times while maintaining a negligible impact on sensing accuracy.

Deep learning-based object detection is widely used in urban sensing, enabling tasks such as pedestrian, pothole, and waste detection. Automotive sensing with dashcams facilitates large-scale, real-time detection across urban environments. However, existing studies primarily focus on detection without estimating object size, which is crucial for event classification. Conventional size estimation methods rely on RGB-D cameras, multiple cameras, or LIDAR, making them unsuitable for large-scale automotive sensing with single RGB dashcams. Monocular depth estimation provides relative depth but does not yield absolute size measurements. To address this limitation, we propose a novel approach that combines monocular depth estimation with a reference object of known size. By comparing the detected object’s pixel dimensions with those of the reference object, its physical size can be estimated. To validate our approach, we developed an automotive sensing platform that detected and quantified household garbage bags using footage from the rear-view camera of garbage trucks. The truck body serves as the reference object, ensuring reliable size estimation. Experiments conducted with real-world data collected using an NVIDIA Jetson TX2 demonstrated the effectiveness of our method. The proposed approach achieves size estimation accuracy with mean squared errors (MSEs) of 20.02 for width and 18.68 for height while maintaining an end-to-end processing rate of 19.21 frames per second (FPS) for detection, tracking, and size estimation.

This research presents a real-Time automotive sensing system for the data of urban garbage disposal. The proposed solution is implemented on an edge computing device mounted on garbage truck where a deep learning based image processing algorithm is implemented to automatically counted the number of collected garbage bags from garbage collection video. A MQTT-based data server was developed to enable data publication from sensing device to server and data accumulation and to facilitate application development. Our system has the functions of high concurrency and low transmission delay, offline reconnection, breakpoint transmission and client authentication. This work is to provide a real-Time, low-cost, reliable and replicable system for the implementation of a widespread sensing network for automotive edge computing and smart city applications.

To cost-effectively acquire people flow data, the use of bird's-eye view cameras that capture wide areas from a distance is being considered. However, videos captured by such cameras frequently exhibit dynamic occlusions, such as those caused by interactions among individuals, and static occlusions, such as those caused by fixed obstacles, leading to a decline in tracking performance. Therefore, this study develops a system to estimate dynamic and static occlusion risk using only detection and tracking information, proposing a general multi-object tracking (MOT) performance improvement method by dynamically controlling tracker parameters based on this risk. Comparative experiments against a baseline method demonstrated that our approach outperformed the baseline on most metrics, suggesting that our method contributes to enhancing tracking performance in occlusion-prone areas. However, an increase in ID switches (IDS) and a decrease in processing speed occurred. The increase in IDS is considered to be due to the unnecessary retention of tracks that should have been deleted due to lifespan extension. This result implies a trade-off between tracking continuity and the consistency of ID assignment.