Requirements for perception

Last updated Jun 24, 2022

• Type atom
• Status needs revision
• On autonomous driving

We need to take into account two things in order to make sense of our surrounding environment and ourselves:

1. Identification: What is the object?
2. Motion: How/Where is it moving?

Unlike humans, computers have difficulty in predicting or figuring out patterns. If we’re able to predict effectively, e.g. how a moving vehicle will behave/turn etc, then we can take informed decisions based on those predictions and improve the driving experience and safety.

Following are some of the elements we need to be able to identify for the task of perception:

1. Static Objects
   • On-road:
     • Lane-markings and road
     • Construction signs
     • Obstructions
   • Off-road:
     • Curbs
     • Traffic Lights
     • Road signs
2. Dynamic Objects
   • On-road:
     • Vehicles
       • 4 Wheelers (cars, trucks…)
       • 2 Wheelers (Bicycles, Motorbikes)
         • These are harder to predict as they have more flexibility and agility.
     • Pedestrians
       • Their movement is much more erratic as compared to vehicles.
3. Ego Localization
   We also need to take into account, our own position in order to construct a full-picture. Sensors like GPS, Odometer, etc are used to measure:
   • Position
   • Velocity, acceleration
   • Orientation, angular motion

Even though a lot of research is being done in this area for the last decade, there are still a lot of challenges remaining.

1. Robust detection and segmentation
2. Sensor uncertainty
3. Occlusion, reflection
4. Illumination, lens flare
5. Weather, precipitation