Autonomous vehicles
effectively ‘remember’ how it dealt with a previous situation and simply repeat its manoeuvre; negating the need for the algorithms to process all the data and find the optimal solution. In the same way that a signal of a reflex reaction passes through the spinal cord rather than the brain, this ‘memory bank of successful solutions’ could reduce the ‘thinking time’ of the vehicle and therefore reduce the time taken for evasive action, possibly saving lives ( How machine-learning algorithms 2018).
Deep neural networks
As the previous section suggested, algorithms rarely work alone and instead, they are grouped and linked together to minimize error and decrease uncertainty in predictions or calculations. A collection of algorithms is called a deep neural network (DNN) and each DNN is dedicated to one task. Pathfinder DNNs help the car figure out where it is going and plans a safe route ahead using a variety of algorithms: OpenRoadNet identifies drivable space around the vehicle, PathNet highlights driveable paths ahead of the vehicle (regardless of lanes), LaneNet identifies lane lines and other markers that can be used to dictate the car’s path, MapNet also identifies lanes and landmarks that can be used to update high resolution maps, and PikoNet predicts the driving centre paths (a line which the vehicle follows) based on trajectories driven by human drivers. The figure shown below illustrates how the Pathfinder DNNs function using a front-facing camera. The arrows in the middle of each lane are produced by the PikoNet algorithms, while the lanes highlighted in different colours are a result of the PathNet and LaneNet algorithms. All these different projections and interpretations help the autonomous vehicle to navigate its way down roads safely and efficiently.
How Do Self-Driving Cars Make Decisions? Burke 2019
Object
detection
and
classification
DNNs detect potential obstacles, as well as road infrastructure such
as traffic lights and signs: DriveNet identifies and perceives other cars, pedestrians, traffic lights, and signs, but does not register the colour of the lights nor the type of sign; LightNet classifies the colour of the traffic lights; SignNet recognizes the type of sign, and WaitNet detects areas or conditions where the vehicle must stop and wait. DNNs also can also monitor the status of sensors and other components of the autonomous system. They can additionally facilitate parking manoeuvres with the use of algorithms such as: ClearSightNet, which monitors how well the cameras can see, detecting limiting conditions such as rain, fog, and direct sunlight, while ParkNet identifies spots available for parking (Burke 2019).
To understand the decision-making process of an autonomous car, we can summarize the process by separating different phases of the system:
1. Hardware and sensors observe the surroundings using various means of technology. They then feed the data gathered from the immediate environment to the onboard computer.
20
Made with FlippingBook interactive PDF creator