���˴�Ƭ

An autonomous vehicle is capable of sensing its environment and moving with little (partially-autonomous) or no human input (fully-autonomous). Machine learning forms the foundation of autonomous vehicles. 

Autonomous vehicles use input from a variety of sensors to “see” the world around them. Sensors may include a combination of LiDAR, stereo-vision cameras, high-definition cameras, radar, internal measurement units, GPS receivers and audio sensors. Using input from these sensors, the autonomous vehicle creates and maintains an internal 3D map of its surroundings. The vehicle uses this 3D map to detect and classify objects on the road such as other vehicles, street signs, traffic lights and lane markings. 

Predictive algorithms are used to anticipate the behaviour of the other objects on the road based on their classification. Route planning software processes inputs from the vehicle’s 3D map and predictive algorithms to plot a path for the vehicle. Instructions are sent to the vehicle’s actuators to control the vehicle’s acceleration, braking and steering. Reactive algorithms are used to respond to unexpected behaviours and emergency situations. This process operates in a continuous cycle, known as an OODA loop (observe, orient, decide, act).

Long-term advantages of autonomous vehicles include:

• Improved safety - With the potential for human error removed, self-driving cars will reduce the incidence of accidents caused by driver error, drunk/drug-driving, fatigue or inattentive drivers.
• Increased work/free time - With humans no longer involved in driving, commuters will be able to use their commute-time for leisure or to undertake productive work, with potential benefits ranging from improved mental wellbeing to boosting the economy.
• Reduced labour costs – There will be no need for drivers for conducting long-haul transport (trucks and trains), public transport, local/delivery drivers, mining and agriculture.
• Improved traffic flow - Autonomous vehicles will integrate with centralised traffic management systems to access up-to-the-minute data on traffic and road conditions to determine the fastest and most efficient route to their destination.
• Improved efficiency – Selection of the most efficient route will result in fuel savings, cutting costs and reducing pollution.
• Increase transport options for people who cannot drive, including the elderly and people with disabilities.

AIML has an established track-record of applying machine learning techniques to the navigation of autonomous vehicles, specifically Simultaneous Localisation and Mapping (SLAM).

Autonomous vehicles are currently being trialled and are entering service around the world in controlled environments and predetermined, fixed routes (e.g. on purpose-built train lines, in public transport, where vehicle routes are predetermined, in mining and farming).

Autonomous trains that run on purpose-built lines are in service around the world. Australia’s first autonomous commuter train (the Metro North-West line in Sydney) entered service on 26 May 2019, while Rio-Tinto operates a driverless freight train that transports iron-ore across the Pilbara region of Western Australia. Likewise, autonomous trucks are being used for ore haulage in mining and trials of driverless buses are underway in cities and towns in Australia, including Adelaide. 

Autonomous cars are perhaps the ultimate application of autonomous technology. Autonomous cars operate on the public road and need to be able to select their route from starting and destination points and their path around obstacles. Likewise, roads vary widely in quality, markings and signage, while geo-location, mapping-data accuracy and internet connectivity has implications on the type and robustness of the autonomy needed.

There are multiple trials currently underway around the world to examine the interactions between autonomous cars, infrastructure and other road users. Legal and ethical issues surrounding the safety of autonomous cars may limit their application in the immediate future.

Traffic is the movement of vehicles on public roads. Transport efficiency is greatly affected by traffic. Traffic can become congested and jammed during periods of peak usage, where roadways narrow or when obstacles (such as double-parked cars, breakdowns, accidents and lane closures) block the road. 

Technologies are available to provide data about the movement of vehicles on the road. These technologies include: inductive loops; pneumatics tubes; IR; audio; Bluetooth and wi-fi detectors; and camera systems. 

Computer vision systems powered by machine learning can analyse video footage collected by camera systems to:

• Detect and classify road users (including pedestrians) 
• Collect key traffic metrics (traffic density, vehicle speed, congestion level, travel times)
• Detect incidents such as breakdowns or accidents
• Detect non-compliant road users (such as unregistered and uninsured vehicles).

Machine learning can be used to collate the data from road sensors and the output from computer vision systems to:

• Analyse traffic data 
• Model the movement of vehicles and the distribution of traffic through a system of roads
• Track the movement of specific vehicles (e.g. heavy vehicles) as they move through a city
• Control traffic lights to optimise traffic flow through specific intersections and along arterial roads
• Detect incidents such as breakdowns or accidents
• Collect data during incidents and special events
• Reroute traffic around incidents (using computer-controlled signalling, road signs and message boards and mobile apps)
• Predict traffic jams.

AIML is currently working with the South Australian Department of Planning, Transport and Infrastructure to develop machine learning techniques for use in traffic management.

Reducing the time, fuel, and labour that is needed to get a product to the consumer can provide transportation companies with a significant competitive advantage.

For a one-off delivery, machine learning can be used to collate data on weather, traffic and road conditions to find the fastest route to a destination.

For a delivery truck, which makes a series of delivery stops in a day, machine learning techniques can be used to optimise the delivery truck’s route; maximising the number of deliveries that can be made in a day. 

For transport companies, who manage a fleet of delivery vehicles that each make multiple deliveries in a day, machine learning can be used to help operations managers:

• Optimise fleet-wide dispatching activities
• Analyse and improve fleet activities 
• Quickly adjust deliveries when unplanned events (such as traffic jams or vehicle breakdowns) occur.

By including an ability to incorporate new data over time (e.g. the actual performance from past deliveries), the machine learning algorithm can adjust its predictions and decision-making based on the new data and experiences.

Machine learning can also be applied to fleet maintenance activities. Machine learning can be incorporated into predictive maintenance systems that use multiple and varied data sources to monitor the performance of the vehicles in the fleet. Using machine learning techniques, it is possible to detect when the vehicle’s performance falls outside of the norm. This allows service managers to plan maintenance activities and ensure that repairs are performed before failure occurs and only when needed, resulting in less downtime and lower operational costs. 

A driver-training simulator is a computer-based training system that allows the user to experience a life-like facsimile of driving an actual vehicle in a safe environment. Driver-training simulators provide a means for assessing driver actions in response to simulated driving situations. Driver training simulators are used for training bus, train and tram drivers and the operators of heavy, mining, military and construction vehicles.

A driver-training simulator consists of a:

• Dynamic simulation model 
• Vehicle emulator.

The dynamic simulation model operates similar to a computer game and uses computationally generated 3D content to represent the environment around the vehicle and a well-defined representation of the function and behaviour of the vehicle that is being simulated.

The vehicle emulator consists of representative vehicular controls and a visual interface allowing the driver to interact with the dynamic simulation model. 

Machine learning can be used to generate 3D content for use in the dynamic simulation model and to improve the fidelity of the physical and visual components of the simulation.

AIML has worked with  to develop machine learning techniques for generating 3D content for use in the Sydac driver training simulators.