The Silicon Trend Tech Bulletin

Icon Collap
...
Home / AI / ML / AI to Predict Others' Behavior on the Road

AI to Predict Others' Behavior on the Road

Published Sat, Apr 23 2022 03:46 am
by The Silicon Trend

 

IMG_20220423_084836_086

 

AI to Predict Others' Behavior on the Road

We, humans, might be one of the most significant road obstacles keeping fully autonomous cars off city streets. If a robot has to navigate a vehicle safely through Boston, it must anticipate what nearby pedestrians, drivers, or cyclists are going for next. Foreseeing others' behavior on the road is a daunting issue. However, the current AI solutions are either too simple, conservative or only predict one agent's subsequent movements.

MIT researchers have developed a deceptively straightforward solution to this challenging task. They classified a multiagent behavior forecast issue into smaller pieces and handled each one individually, so a computer could solve it in real-time. The forecasting framework first predicts the relationships between two road users, i.e., which pedestrian, vehicle, or cyclist has the right of way, and which agent will yield. Then, by leveraging those relationships, the future paths for myriad agents can be predicted.

These estimated paths were more precise than those from other ML models compared to actual traffic flow in a large dataset compiled by robotic driving firm Waymo. The MIT method even outperformed the company's recently introduced model, and as the researchers classified the issues into small pieces, their approach used minimum memory.

 

Also Read: Why it's Difficult to Create AI Unbiased and Fair?

 

Myriad Small Models

The researchers' ML approach, called M2I, takes two inputs: path paths of the pedestrians, cars, and cyclists interacting in a traffic setting like a 4-way intersection, and a map with lane configurations, street locations, etc. Using this data, a relation forecaster infers which of two agents has the right of wat first, categorizing one as a passer and the other as a yielder. Then a marginal predictor guesses the path for the passing agent; since it behaves independently.

Another prediction model is a conditional predictor that guesses what the yielding agent will do based on the passing agent's actions. The system forecasts some different paths for the passer and the yielder, computes the probability of each one individually, and then chooses the six joint outcomes with the highest possibility of occurring.

 

Also Read: Meta Rushes to Release First AR Glasses in 2024

 

Real-life Driving Tests

The MIT researchers trained the models using the Waymo Open Motion Dataset, which contains several real-traffic scenes involving cyclists, cars, and pedestrians recorded by LIDAR sensors and cameras on the firm's robotic cars. To determine the precision, they compared six prediction samples of each approach, weighted by their confidence levels to the absolute paths followed by the others in a scene.

An advantage of M2I is it breaks the issue into smaller pieces, making it easy for the user to understand the model's decision-making. This might help users put more trust in robotic vehicles in the long run. However, the system can't account for situations where two agents mutually influence one another, like when two cars each nudge forward at a four-way stop as the drivers aren't sure who will yield.

 

Also Read: New Health-tracking Features and Improved Notification for iPhone Users