AI to address road traffic accidents

Road traffic accidents is a fatal scourge in the context of Bangladesh. The question is, whether AI can work to prevent road traffic crashes through predictive analysis.

Sadhli Roomy

AI to address road traffic accidents in Bangladesh.

Can real-time monitoring and automation make a dent in one of the worst problem in Bangladesh and globally?

Infographic sourced from The Daily Star.

If we combine the reported fatality cases from 2018 and 2019 alone, equaling 9,666 deaths, road traffic accident becomes a greater threat than COVID 19 in Bangladesh. Nirapad Sarak Chai corroborated the data from two separate reports provided by the police and BUET's Accident Research Institute (ARI)on which they counted 20 percent of injuries as deaths according to international standards.

However, there is no way to put an actual number on the deaths caused by road traffic due to many reports not showing up to the police. An example would be year 2013, where the Government reported 3,296 deaths according to BBC which came into conflict with a major WHO report in 2015 which used modelling to estimate that there were 17,349-to-25,283 deaths in Bangladesh. Hence it is fair to assume that the actual number of road traffic deaths are higher than reported cases.

The risk factors of road traffic are very well defined; inclusive of human errors, unsafe infrastructure, unsafe vehicles, poor enforcement of laws, and lack of post-crash trauma care. There is no concrete data to support this but 'human errors' are possibly the most notorious concern, buses, trucks, and covered vans are the most notorious vehicle types to be involved in an accident, and pedestrians are the most vulnerable group according to a report by ARI.

What are 'human errors' in road traffic accidents?

From technical terms, we agree with Green and Sender's (2013) definition of human errors - 'mistakes' a driver makes when they are on the road due to impedance of perception, attention and memory. A driver processes vast amounts of visual information on the road such as the road, vehicles, people, signs, the environment among others. There are other mental processes at work as well such as auditory information and recalling memories - all of which demands 'attention'. When the capacity for attention exceeds a driver's capacity, human errors are likely to incur and increase probability of accidents. This typically happens when:

  • drivers are speeding
  • visibility is compromised
  • they are more focused on a particular object
  • the capacity of the driver is lowered through fatigue or substances

Now all of these avenues can have solutions of their own and many of them do but there has been little work done on addressing the affect influenced by physiological (bodily) and psychological (emotional) input of drivers.

To put this into perspective, lets take Khurshid Alam as an example, a 45 years old private bus driver plying his trade at the heart of Bangladesh's capital, Dhaka. Alam drives a bus for three days a week, each time pulling a 17-hour shift.

"I start driving at 6 am and continue up to 11 pm..."

And in doing so, he gets to earn around BDT 1,200/- per day - roughly $14. Alam is not salaried like most of the other drivers and earns his keep by commission and trips he takes; which again leads to bus drivers racing each other to pick up passengers. Severely overworked, underpaid, fatigued, abused by passengers, and having to endure the heat during the warmer months, drivers like Alam tends to be always on the edge.

A coping mechanic is to use drugs or engage in substance abuse; evidenced by over half of bus drivers having reported to use them. Quoting from the same article, a driver said:

“Most of us drive vehicles for 12 to 16 hours a day though the labour law limits the timeframe to eight hours...”

In summary, all of the factors mentioned to fit into a driver's mental capacity above which trigger accidents, seem to be present in that of heavy vehicle drivers. The day I started writing this piece, a bus and a truck took the lives of 13 people in 2 districts of Bangladesh.

Traditional countermeasures

So how might we combat this? Traditionally, public awareness campaigns for both drivers and pedestrians, stricter enforcement of law, infrastructural improvements, and one-to-one counselling of drivers follow suit in order to minimize road traffic accidents; and it has been going on for quite some time. While effective 'when' provided, we have to understand that most of these interventions are dependent on human memory - a fleeting and unreliable thing if the said countermeasure is not reinforced continuously which can essentially be borderline impossible to do at scale.

Use-case of AI to mitigate road traffic crashes: Will it work?

Automation and AI are beautiful things and there are many novel ideas and executions happening across the world. Imagine if there was EasyMile's driverless passenger shuttles or a fleet-based rendition of it available in Bangladesh, completely overhauling the urban and regional transit systems. One could argue that accidents could be minimized in event the human element is removed. However, just like EasyMile's constant battle with labor unions, we can expect a similar, albeit more aggressive, response as soon as it is deployed.

Even if we have something like comma two for buses and trucks, we could ideally see a fair bit of development and mapping time involved for deploying the solution in a region like Bangladesh - even in a what-if scenario.

So 'no', solutions made for developed economies, such that of the US, UK, Germany, the EU, with superb urban and regional infrastructure may not replicate the same use-case in Bangladesh since the enabling environment is fundamentally different.

So what of can be do to ensure scalability and contextual appropriateness of deployment?

Considering contextually appropriate solutions, we found Prof. Hajar Mousannif's and her team's work on the Collision Avoidance System (CAS) - a real-time crash prediction system, to be the most appropriate solution that takes a crack at integrating AI to extract information of a driver's psychological and physiological states enabling others to act on it. It's quite possibly the only or one of the most comprehensive frameworks in the market today.

Prof. Mousannif is a Professor of Computer Science (Informatique), Cadi Ayyad University, Marrakesh, Morocco, serves as the coordinator of their Master programme in Data Science, and, as of recent named as the golden winner of the WomenTech Global AI Inclusion Award 2020.

Watch her keynote at the 7th ICMCS 2020 as she explains CAS, the driver behavior analysis model they curated, how her team is working to mitigate imbalanced data that are inherent to crash-related observations, and the machine learning fusion framework for CAS.

Originally published by the MSTI Events YouTube Channel on 02 October 2020.

The things that caught my eye are the driver behavior analysis framework and the information fusion avenues - taking in input from multiple devices to articulate accident probability and harnessing a form of early warning system. Although early in its development, CAS - paired with other solutions - can provide a gateway to address driver's physiological states (and hopefully their psychological states in the future) - something that is missing in the solution ecosystem in Bangladesh and possibly throughout similar countries like Pakistan and India. Not to mention the CAS fusion framework also accounts for driving inputs and vehicle kinematics to draw up a range of factors that attributes to road traffic accident risks. Following is an image from Dr Mousannif's keynote which shows the simulation her team undertook to test out their fusion framework:

A small use-case can be using a pilot fleet of vehicles, integrating the CAS system, and feeding this data to Traumalink - a social enterprise that works to build volunteer first-responders to road traffic accidents in Bangladesh who can then use this data to strategically place their volunteers in specific geographies. Another use-case is to provide alerts to drivers when they exceed a threshold of fatigue and/or other physical conditions, detected through input devices that tracks their facial expressions, pulse, throttling pattern, and kinematics. They are given voice-activated reminders on what they should do to avoid accidents added with other nudge tactics if possible.

A larger use-case is to embed policies to force transport owners to install CAS modules and integrate a complete system with Bangladesh Road and Transport Authority's (BRTA) Road Safety Cell to monitor predictive data to mobilize resources in critical areas and make strategic decisions on minimizing road traffic accidents.

Yes, there is still a ways to go before likely solution is developed but that future isn't too far away. The sheer potential of the ways CAS can play out in the spectrum of innovation is thrilling, along with its relative appropriateness of deployment in the global south where AI can play a critical role in catching and putting one of the worst killers in human history behind bars.

There are about 400,000 registered trucks in Bangladesh presently, of which 100,000 runs on the road everyday.

Dhaka Tribune (2020)

Number of registered buses are 44,374 according in 2018.

BRTA (2018)

Close to a million drivers do not have a license.

Dhaka Tribune (2019)

Let's start a project together.

Get high-quality and pixel-accurate labelled data through us. We bring the best-in-class, scalable, and adaptive annotation and quality assurance services. Begin the final trek towards fine-tuning your supervised learning model with us.