The number of car casualties every year > the number of casualties in WWI

Almost 3 years ago I was in a car accident where the car flipped over 4 times. It’s been 1030 days and I can still run the entire accident in my head. I was told if I were not in an SUV, I would have been carried out of that car. To say the least, it was definitely a memorable experience. The reason?

A distracted driver in a pickup truck in the left lane.

I’ve been in 4 other accidents in my life, 3 of which were caused by distracted driving and I’m definitely not the only one experiencing this. To give you some numbers, car accidents are the cause of around 50 million casualties every year, which is more than the number of casualties in WWI. In fact, car accidents are the leading cause of death of young people ages 10–29.

This is a problem.

27% of these road accident deaths are also caused by distracted driving. Another 12% are caused by drowsiness.

To put that into perspective, that’s almost half a million people dying every year.

Every single one of those half a million people have their own story. A story of death because someone was overconfident and careless.

The Problem

Most people actually admit to texting, eating, and putting on make-up while driving, yet, they continue to practice these actions.

Why?

As explained by psychologists, the reason why people continue to do bad actions, even though we know they’re bad is because of a concept called Reinforcement Theory. In essence, psychologists say that if nothing there is no consequences to our actions, we continue to do engage in it, with minimal fear of repercussion.

Another problem that prompts humans to believe that we can drive while distracted, is multitasking: an illusion that we created for ourselves. Although we’ve convinced ourselves that this is our ability to do more than one action at once, multitasking is actually the brain’s ability to rapidly switch from one task to the next. In fact, a study conducted by the University Of Utah proves that the reaction time of a person texting while driving is just as bad as that, of a drunk person. The study also showed that 80% of car accidents that are caused by drunk drivers occur between 6 p.m. and 6 a.m. because the drivers are fatigued. However, people fail to believe that it is impossible to multitask and continue to claim that their brain can handle doing two things at once.

Inattention blindness

This concept basically explains that even when looking at something directly, when we are not paying attention, our brain does not process it. This correlates to when drivers look directly at road conditions but don’t really see them because they are distracted by a cell phone conversation or by their lipstick in the mirror.

Because humans are so confident in themselves, it’s extremely difficult to convince them otherwise. If we really want to prevent distracted driving, then there has to be a reason. Why should I stop texting? How do I benefit from such an action? There needs to be an incentive.

Our Solution — The Technical Aspect

Our solution to combat this problem will use an Artificial Intelligence algorithm to analyse and detect distracted driving based on several variables and then, perform several actions with this data.

The way we’re achieving this is through a Deep Learning neural network. Deep learning is a method of how we can mimic the human brain via a neural network. It’s a subset of Machine Learning that uses an algorithm inspired by the brain. The ideology of neurons is implemented into neural networks with a neuron in a neural network being known as a node/perceptron.

In a neuron, the dendrite receives signals from other neurons. This mainly includes electric impulses that are sent to the neuron from other neurons. The cell body then helps sum all the inputs (which are the electrical impulses that contain a letter, value, or symbol for example) to get a final value. From there, an output is given which is then released to other neurons by the axon. This is literally how neurons in a neural network work. Each perceptron receives multiple inputs where it is past data or image pixels, applies various transformations (algorithms) and then provides the output.

This is exactly how a neural network works. we do a weighted sum of numerical values instead of electrical impulses.

This is what a basic neural network looks like

The input layer would be our input data. From there, we’d perform the weighted sum (Wx+b) of the inputs with the weights and biases and take that value into each hidden layer. This gets repeated until we get an output.

We also use something known as an activation function. An activation function is a function that helps produce a tangible output based on the weighted outputs. We’d take our weighted outputs, run them through the activation function, and then input that output into the next node in the hidden/output layers.

The outputs are normalized as seen above. The overall significance of activation functions is that it helps decide what should be outputted to the next perceptron. It helps the model learn complex patterns in the data by converted the weighted sum into some form that can be taken as an input to the next node. Activation functions also help keep the output value restricted to a certain limit. By normalizing all the outputs, the model won’t have to face computational errors causing it to break. It also helps us solve the vanishing gradient problem. What this means is that gradient descent uses an algorithm known as backpropagation. When utilizing high-level neural networks, the gradient tends to vanish due to the depth of the network and the gradients shifting to zero. Activation functions help solve this by not shifting the value to 0.

The way we detect distracted driving is through the use of a Convolutional Neural Network (CNN). CNNs are a type of deep learning algorithm that specializes in image and object recognition. The overall goal of CNNs is to reduce the amount of “richness” the image has into a form that’s extremely easy to process while maintaining the main features and details required to produce a proper prediction.

We use certain techniques such as Convolutions, MaxPooling, and Dropout to help the model understand the data extremely well along with not overfitting along with Data Augmentation to help ensure that the model will have extremely high accuracy. At Undistracted, our state of the art Convolutional Neural Network has an accuracy of over 99%.

The Actual Product

The algorithm will be implemented into an app that is connected to a webcam in the right-hand corner of the car.

The app will constantly be analysing images from the webcam to detect when the driver is distracted, by using the features it previously established in each one of the actions. All data will also be stored locally unless the driver consents for us to use it for model training. Therefore, there is no way of any of this data to be used against them.

As the driver is driving, the app will alert them when they need to pay more attention, similarly to the seatbelt alert in current vehicles.

In continuation, if improving our model, it will also take in external factors, such as road conditions, location, city (for specific laws), time of day, number of cars around etc. to accurately determine distracted driving.

However, an alert system may not convey this idea enough so…

Why would people want this?

A major incentive for people to buy or install our product is to reduce their insurance premiums. In fact, several insurance companies have already implemented a similar program into their premium calculation algorithm.

For instance, TD Insurance has created an app called, “MyInsurance”, used by 70% of their clients, which tracks speed, acceleration, braking, cornering and the time of day while a client is driving. This company reduces premiums by 5% for only participating in the program and after analysing the yearly score and progress, can reduce premiums up to 25%, which is an equivalent of about $400 yearly for an average driver.

Since distracted driving is a significant issue and the reason for many insurance claims, companies will increase premiums by 15–24% of their client receives a distracted driving ticket in Ontario. Hence, it’s evident that they do care and will be interested in our proposal as now there’s an incentive to save money.

Varying from person to person, and whether their motivation for this product is internal (would like roads to be safer) or external (would like to decrease their premiums), another incentive that arises is accountability while driving. For example, several statistics show that a good fraction of tired or distracted drivers are oblivious to the fact that they are creating an unsafe situation. However, with evidence of them doing so, drivers may be more prone to pay attention to their actions at the wheel. And for people who are internally motivated to prevent unsafe situations and hold themselves accountable, this is the perfect solution.

How do we know this can be implemented?

The TD app is used by 70% of their clients and the other 30%, being people that are not very good with technology or are skeptical about where this collected data is going and whether it will be used against them. TD also states that it is illegal for that data to be used for anything other than the calculation of premiums.

In fact, an agent from TD Insurance claims that “The idea of detecting distracted driving is definitely implementable into our company and can be put under review to improve our MyAdvantage App.” When asked, whether people would use this feature, he said “Although it would take a few months for people to get used to the idea, people value money over everything many other aspects, especially when their insurance costs $7,000 per year.”

Why this?

The 8th leading cause of death is car accidents and other than the development of self-driving cars, not much progress is being done to prevent it. By alerting the driver that they are distracted, as well as having the potential possibility of alerting others, this risk can be irradicated.

Why us?

Because of a desire to truly create impact and save lives, we have created a unique solution, in comparison to other companies.

1. Better accuracy in the AI model

With a preliminary dataset, we have achieved a 99.5% accuracy rate in the algorithm. As we build more data and continue training our model, this accuracy has the potential to improve up to 99.99% accuracy.

2. Insurance Incentive

We will be working with insurance to decrease premiums for clients based on their scores in the app. However, their premiums are also impossible to increase, since the client chooses to send any and all data.

3. Guaranteed Privacy

We guarantee that all data is stored locally and only the client’s score is sent to their specific insurance company.

How will we improve our product?

1. We take into consideration outside factors

As a next step, our model will consider outside factors, including by-laws, such as the Mississauga by-law of: no drinking water while driving. It will also consider the safety situation in the surroundings of the car, including how many cars are around, what speed the driver is driving at, if there are people around and more.

2. Think about privacy

100% of the data will be stored locally and if the user agrees, some of the data can be used for training, insurance and monthly conclusions about the driver.

3. Work with car manufacturers.

A potential next step is to also begin working with car manufactures to implement a webcam into the structure and OS of the car.

Why now?

Although self-driving cars seem to be the centre of innovation, according to TechRepublic, Forbes, Technative.io only 10% of cars will be automated by 2030. It may seem like we are on the verge of automating cars, however, we are still very many steps away from doing so. Therefore, our solution will still be viable for decades before self-driving cars will reach at least 30–50% of the cars on roads.

And therefore, change needs to happen now. I won’t settle for that change happening in 10 years. I don’t want to be in a 5th car accident. So with undistracted, let’s make that change now.