Brian Fuller President,
Sensys Networks
There are several key reasons why Al has moved to the forefront of ITS innovation, but a few seem to be the driving forces of this change.
There are several key reasons why Al has moved to the forefront of ITS innovation, but a few seem to be the driving forces of this change. First and foremost, advance- ments in the computational capacity of processors have enabled more complex Al models like those built with deep learn- ing. This has also enabled sophisticated applications to run in edge-processing environments where central servers are not required for real-time applications. Tied to processing-power increases are advancements in deep-learning models within Al. Deep learning has significantly improved the performance of Al while reducing the complexity of implementa- tion. For example, it has dramatically reduced the issue of false positives in video analytics. Prior to deep learning, false positives created poor results for many video analytics applications in ITS. Utilising properly-trained deep-learning models can reduce false positives by an order of magnitude. Deep learning also moves the product offerings closer to true automation, versus all of the handholding required to support earlier traffic Al systems. Sensys’ sister company, Citilog, has been developing machine-vision Al technology for 28 years and has deployed it on over 60,000 cameras through more than 1,600 implementations worldwide. In 2019, it made a major step up to deep learning by integrating its vast, real-world dataset into DL models. This seems to be the key element that many newcomers to this space are missing.
Al will not be the perfect fit for all traffic applications, but it can be combined with other technologies to reach new heights.
Everyone has an algorithm, but do they have the right comprehensive dataset for their environment that fully optimises their algorithm as they train their Al model? More generally, the evolution of machine vision in other larger markets and industries has enabled the advancement of these technologies in the transportation industry. At the same time, a knock-on effect of the prevalence of AI in these other markets – especially in consumer applications has led to more accelerating both interest and adoption. Within the traffic space, graphical user interface improvements, including those enabled by the high-quality video now readily available, have made these tools both easier to interpret and more valuable. Some common examples in traffic Al are bounding boxes and paths of travel that highlight different modes of traffic as well as critical traffic incidents. Al will not be the perfect fit for all traffic applications, but it can be combined with other technologies to reach new heights. Using detection as an example, radar is still ideal for highway data collection because of the low cost, accuracy and simplicity. And wireless sensors work where other detection doesn’t, because they can be placed far from power and networking sources. Recognising this, Sensys Networks recently launched a solution named MultiSens Intersection, which combines video Al at the intersec- tion with wireless sensors for midblock or advance detection.
Al is excellent for differentiating vehicles and vulnerable road users, and edge processing can be leveraged to combine the detections from the cameras and the wireless sensors. This solution can actuate traffic signals in real-time and provide higher-level metrics using better input data than video alone.
