Cloud Chaser: Real Time Deep Learning Computer Vision on Low Computing Power Devices (1810.01069v2)

Abstract: Internet of Things(IoT) devices, mobile phones, and robotic systems are often denied the power of deep learning algorithms due to their limited computing power. However, to provide time-critical services such as emergency response, home assistance, surveillance, etc, these devices often need real-time analysis of their camera data. This paper strives to offer a viable approach to integrate high-performance deep learning-based computer vision algorithms with low-resource and low-power devices by leveraging the computing power of the cloud. By offloading the computation work to the cloud, no dedicated hardware is needed to enable deep neural networks on existing low computing power devices. A Raspberry Pi based robot, Cloud Chaser, is built to demonstrate the power of using cloud computing to perform real-time vision tasks. Furthermore, to reduce latency and improve real-time performance, compression algorithms are proposed and evaluated for streaming real-time video frames to the cloud.

• The paper introduces a novel cloud-based architecture that enables real-time deep learning object detection on resource-constrained devices.
• The paper employs innovative latency reduction techniques, including image compression and selective region transmission, to achieve processing delays between 494 ms and 756 ms.
• The paper demonstrates robust empirical results on a Raspberry Pi robot using the YOLO framework, confirming the system's practical effectiveness.

An Assessment of Cloud-Based Real-Time Computer Vision for Low-Power Devices

The paper presents a compelling solution for enabling real-time deep learning-based computer vision tasks on devices with limited computational resources, such as IoT devices and mobile phones, by leveraging the capabilities of cloud computing. The authors introduce a system design termed "Cloud Chaser," utilizing cloud resources to perform computationally intensive processes remotely, thereby negating the need for dedicated high-performance hardware on the local device itself.

Contributions

The paper details three primary contributions to the field:

1. Software Architecture Design: The authors propose a novel architecture employing cloud resources to perform real-time deep learning object detection. The architecture uses asynchronous communication and multi-threading to handle real-time data transmission between cloud instances and local devices efficiently.
2. Latency Reduction Techniques: To mitigate the inherent latency issues involved in transferring data to and from the cloud, the authors design and evaluate compression techniques. These approaches include reducing image resolution, applying image blurring, and selectively transmitting only the regions of interest within an image frame.
3. Empirical Evaluation: Through experiments conducted on a Raspberry Pi-based robot, the authors evaluate the proposed architecture under practical scenarios. The robot demonstrates real-time object detection capabilities, backed by cloud support, establishing the feasibility and effectiveness of the system.

Key Findings

The authors detail several key numerical results from experiments:

• The latency from local device image capture to cloud-based processing and return was shown to average between 494 ms and 756 ms, depending on the time of day and network congestion levels, using a GPU-equipped cloud server.
• Compression techniques reduced file sizes effectively, with JPEG image size reductions of up to 23% when regions of interest were isolated and transmitted selectively. This facilitated decreases in transmission latency for the object detection application.
• The YOLO (You Only Look Once) framework utilized for detection demonstrated robust performance, maintaining real-time efficacy with minimal compounding latency from cloud processing.

Implications and Future Work

The implications of this research extend to various real-world applications requiring object detection and real-time processing on resource-limited platforms, such as autonomous drones, smart surveillance systems, and mobile robotics in smart environments. The use of cloud infrastructure for such applications demonstrates a cost-effective alternative to deploying high-performance hardware directly on local devices.

From a theoretical standpoint, the work suggests new avenues for exploring communication protocols and data optimization strategies in tandem with sophisticated neural architectures like YOLO. Continued improvements in cloud resource management and further reduction in latency are anticipated through evolving networking technologies and more efficient algorithmic designs.

Speculation on Future Developments

As cloud computing and neural network models continue to evolve, efficiencies are expected to improve significantly. Enhanced bandwidth capabilities with 5G and beyond could substantially drop latencies, while the advent of edge computing could bridge some gap by positioning processing capabilities closer to the end devices. Additionally, the future may see enhanced integration capabilities with increasingly sophisticated sensors, synergizing to provide more seamless and powerful real-world data integration.

This research provides a foundation upon which further explorations into hybrid cloud-edge computing frameworks might build, potentially unlocking even more powerful applications for resource-constrained devices in distributed processing environments.