Machine learning technology is helping to improve traffic monitoring and data analytics in the transportation sector. It enhances detection accuracy and turns the collected data from cameras and sensors into valuable insights.
Machine learning technology is helping to improve traffic monitoring and data analytics in the transportation sector. It enhances detection accuracy and turns the collected data from cameras and sensors into valuable insights. The insights can then be used for event prediction
to prevent road traffic injuries, eliminate peak-time traffic congestion, enhance operational efficiency or for future planning of traffic infrastructures.
Rise of Machine Learning
Machine learning typically deals with large amounts of unstructured data like sounds and video images. Using artificial neural networks to mimic the way the human brain understands information, machine learning-based systems can learn and analyze different patterns and deliver accurate results in various conditions, unlike rule-based systems.
As traffic monitoring centers need to handle a large amount of data in an ever-changing target environment, this seems to be one of the best applications for this technology. Juber Chu, CEO of ACTi
, indicated, “The traffic rules are generally fixed and clear. Therefore, supervised machine learning algorithms are widely applied in this sector.”
Eric Olson, VP of Marketing at PureTech Systems
, said, “The transportation industry is based on a vast set of rules as well as predictable good and bad behaviors, making it a great candidate for the use of intelligent video (and non-video) based automation. In this environment, machine learning has the capability to not only learn rules, but also learn more complicated scenarios and assess the best course of action.”
The technology is relatively new in the field of transportation and is still evolving. “Applying machine learning in this area is still in its early stages. Machine learning is going to become the standard one day. Once people start to see the successful applications, it will grow exponentially,” said Zvika Ashani, CTO of Agent Video Intelligence (Agent Vi
Guy Baron, CTO of Qognify
, expects more and more smart sensors that pack intelligent algorithms to be deployed as part of the transportation infrastructure. These can ingest and process large amounts of data in real time while helping to optimize operations by automatically identifying “cause and effect” and providing recommendations to operators for altering operational decisions to enable optimized outcomes. “This is already happening today. We are likely to see significant growth in this trend as the potential of IoT is realized. Vendors, systems integrators and customers put machine learning, big data and AI as a key point in their solutions and requirements,” he said.
Object Detection and Event Prediction
Machine learning helps enhance detection accuracy in identification and classification like vehicle size classification, compared to other techniques. Constant Rutten, Marketing Applications Video Systems at Bosch Security Systems
, said, it allows for higher fidelity and specificity in the classification of vehicles, for instance, with detection of the number of vehicle axles for tolling or distinction between small and large trucks.
Ashani indicated that three or four years ago, applying video analytic technology to track vehicles and count vehicle flows in a relatively crowded environment (such as a highway or a city street where there are a large number of objects like people, cars, buses, motorcycles, bicycles) was normally not accurate enough. “The addition of machine learning significantly increases accuracy for object classification and recognition,” he said.
The machine learning algorithms analyze data coming from surveillance cameras and sensors, and detect unusual patterns when they occur, like roadblocks or crowds in the middle of the street where they shouldn’t be. “The system learns the patterns around the city at different times of a day, different days of a week, and constantly updates the patterns,” Ashani added.
Deep learning is a branch of machine learning that is popular for image and video processing, and PureTech Systems uses it to complement other video analytic capabilities. Olson said, “Deep learning is very effective at classifying objects and eliminating noise within the video image. When presented with a complicated scene, or a detection scenario using a high-resolution video feed, we have found that deep learning can be extremely useful in accurately identifying targets.”
Chu said, “When the system is trained with enough patterns, it can recognize whether there is a car accident or a gathering of cars.” The algorithms can analyze different angles and distances among cars to make valid judgements. “In addition, we can use algorithms to predict events. For example, when the system detects abnormal traffic flow or behaviors, it can analyze and infer from labeled training data,” he added.
Traffic accidents are unavoidable and hard to predict. Smart prediction is helpful to avoid accidents like collision prediction at intersections or provide actionable insights. Rutten indicated that a partner company of Bosch Security Systems has a machine learning-based prediction program that uses the metadata generated from Bosch cameras to obtain vehicle trajectory data, next to other sensor data coming from radar or in-vehicle GPS-based localization systems communicating through DSRC to do collision prediction. The prediction program runs on a roadside computer unit near the intersection.
It’s important to have predictive capability to know which areas will be affected by accidents. Ashani said, “If there is something happening in one area, a smart system should be able to predict that another area will be affected in a few minutes.”
In the near future, we expect to see the smart systems make decisions themselves and help to deal with emergencies automatically. Chu of ACTi said, “The system can learn traffic accident patterns and how to deal with the incident. When there is a car accident, the system will determine the number of ambulances needed, and connects to the nearest ambulance and police car via GPS to support.”
More Efficient Monitoring
Previously, human operators need to spend hours sitting in front the monitors for viewing lots of traffic videos. However, fatigue and distractions may affect the operators’ responses to accidents.
Anthony Fulgoni, Chief Revenue Officer for Calipsa
, indicated that a machine learning solution can be used to assist monitoring teams. Also, the system could send a heads-up notification when an accident is detected. “The machine learning platforms can provide 24 x 7 x 365 operations for counting, classification, tracking, speed estimations and anomalies identification without getting weary. It means that the operators can be more effective in taking action when incidents happen,” he explained.
Machine learning performs pattern matching and cognitive tasks on large amounts of data that traditionally need to be scanned and processed manually. Baron said, “By using machine learning algorithms that perform object recognition in video footage, we can provide our users a simple and fast search mechanism to analyze scores of video content. It allows them to dramatically reduce the time, effort and ‘cognitive load’ of performing this task.”
Ashani indicated, “The machine learning-based system is learning by itself, and doesn’t require users to do a lot of tuning and tweaking of the parameter in order to achieve high accuracy. This is a very big enhancement. The user doesn’t need to understand so much about how to tune such a system to get high performance because the system can tune itself. For the users and installers, the systems become much easier to use and deploy.”
Baron thinks it will be the norm that systems will automatically identify road hazards, security issues and predict traffic congestions while proactively prescribing corrective action in the near future. “Machine learning and AI will play a central role. Fueled by data pouring in from different sensors and data sources, intelligent algorithms will be automating more and more of operational procedures being carried out today by humans, allowing for risk to be mitigated and enhancing operational efficiency,” he said.
Improved Accuracy and Identification
One of prominent benefits of machine learning is the improvement in the accuracy of analytics algorithms and reliability of video surveillance cameras.
Machine vision technologies have been used in traffic monitoring applications like automatic incident detection systems for over ten years. Rutten indicated that machine learning improves the output quality of these automatic monitoring applications due to the higher accuracy of incident detection and more reliable detections in adverse weather and low-light conditions. “Compared to older machine vision technologies, it also delivers improved tracking of pedestrians, bikes and vehicles in complex traffic scenes with many vehicles and persons moving in front of the camera. It enables new types of analysis that were not practical before, like detecting delivery vehicles that are double parked in inner cities causing severe traffic jams,” said Rutten.
The technology extracts specific characteristics from vehicles and provides better identification capability. Daniel Chau, Overseas Marketing Director for Dahua Technology
, said, “Formerly, video and audio oriented intelligent analysis technologies have been riddled with bottlenecks, are not very precise and cannot be truly used in business situations. In contrast, after adding machine learning models to the system, there is a significant increase of accuracy in the recognition of people, cars, roads and other important aspects of transportation. To begin with, the ability to recognize complex car plate numbers, such as those in India, will be greatly enhanced.”
Dahua started to research on intelligent algorithms in 2009, and is now very close to applying deep learning technology on its cameras for vehicle and human identification and statistical analysis. Furthermore, the machine learning algorithms can identify car features like type, make, model and color in a more systematic way. “Combining various elements in one search, it becomes possible to identify a target vehicle even if the license plate is not captured,” Chau added.
Furthermore, the semiconductor technology has been improving, enabling system on a chip (SoC) to be more powerful and less power hungry. Thus, Chu expects to see more and more surveillance cameras supporting machine learning algorithms as they are being equipped with higher computing capabilities.