Deep learning on the edge
|This post is written by Eric Toffin is the CEO of Citilog. Read more about Eric at the end of this post.|
Increased processing power in the latest ARTPEC chip technology offers new possibilities for fully embedded deep learning applications at the edge. Citilog CEO Eric Toffin examines the evolution of intelligent traffic management solutions and looks ahead at the benefits of edge-based deep learning technology.
The benefits of edge-based deep learning
As a pioneer in intelligent transportation systems and with over 20 years of experience, we at Citilog have experienced countless technology firsts. Two of the most significant for our industry have been the shift from analog to IP video between 2008 – 2012, and the transition from server-based traffic management solutions to cloud- and edge-based solutions beginning in 2016.
Since the introduction of cloud-based networks, IoT adoption has rapidly increased. And experts predict that there will be over 75 billion IoT devices by 2025 – representing a five-fold increase in 10 years. But these devices generate so much data that not even new 5G networks can manage alone.
So the race is on, to de-centralize computation and services from the cloud to the edge of the network, while simultaneously increasing the processing power of edge devices. Although edge-based solutions are not predicted to replace cloud or server-based computing, we found that 70% of new projects in 2019 included edge-based analytics. And there are some definite advantages.
- Lower bandwidth consumption
This is probably the most obvious. Cloud-based training and inference models require devices to transmit huge amounts of raw data to the cloud, thus consuming enormous network bandwidth. Conversely, edge-based deep learning transmits only the analyzed metadata and requires a small fraction of available bandwidth.
- Reduced latency
For high speed traffic incident management, latency delays can make the difference between capturing the root cause of an incident or missing it entirely. Edge-based deep learning now makes it possible to provide real-time applications with no central communication – such as wrong-way driving detection, for which detection time is of the essence.
- Better reliability
As cloud- or server-based analysis rely heavily on wireless networks, any disruption in coverage can have a significant impact on the results. However, with edge-based deep learning, all computing happens on the device itself and is less susceptible to the impact of intermittent outages in network coverage.
- Privacy compliance
Maybe not top of mind, but personal information such as license plate numbers are increasingly protected by privacy legislation. Minimizing the amount of personal data transferred to the cloud can also help ensure privacy compliance.
- Cost savings
Transmitting analyzed metadata rather than large amounts of raw data eliminates the need for additional storage devices or excessive cloud-based storage fees. In addition to saving on hardware and storage costs, the system consumes far less power – minimizing both expense and environmental impact.
Deep learning for better performance
Irrespective of whether your deep learning solution is server-, cloud- or edge-based – or a hybrid version thereof – it’s important to understand where the true value of deep learning lies. After all, video analytics have been used for years to help traffic authorities monitor the overwhelming number of cameras on roadways.
As we discussed in an earlier blog, the biggest challenge with simple alert-based monitoring is the at time high number of false alarms generated. Specifically for traffic management, this means that effective video-based incident management solutions have until now been limited to tunnels with controlled light conditions.
However, the introduction of deep neural networks (DNN) to detect, identify and classify data, now make it possible to differentiate vehicles, bikes and people. Further, video solutions can be trained to recognize shadows, reflection and glare in variable light conditions, thus decreasing false alarm rates and expanding the use of effective video-based incident management to highways and bridges.
Beware the deep learning “hype”
Like many technology trends, artificial intelligence (AI) creates a lot of confusion and false expectations. But the reality is that not all deep learning solutions are created equal. Like any intelligent video application, the quality of real-time traffic management solutions relies on two main factors – a robust dataset and image quality. At Citilog, our solutions are “trained” on tens of thousands of relevant high-quality video images coming from 20+ years of experience in traffic incident management. And it’s evidenced in their performance, reliably detecting actionable events while producing the industry’s lowest false alarm rate. The unique combination of Citilog deep neural networks with Axis video technology now makes edge-based deep learning possible. And our solutions are providing better detection and lower false alarm rates for traffic authorities around the world.
Learn more about our solutions for smarter roads:
|Eric Toffin is the CEO of Citilog. With a background in mechanical engineering and 20 years of field experience, Eric has a strong understanding of both the technology and market requirements for the Intelligent Transport System (ITS) industry. He joined Citilog in the early days of development and helped grow the company into a leader in the industry of video analytics for traffic management.
Through video-based traffic management solutions for smarter mobility, Citilog contributes to roads where traffic moves safely and more freely.