Every Pixel Matters

Now-a-days, vision based collision avoidance systems are very popular. In last few years, DL models has significantly improved the capabilities of cameras with detecting/classifying objects in real time. Stereo cameras those are known for depth estimation and capable to detect and classify road obstacles, are good competitor of Radar systems for forward collision avoidance system. In Stereo Camera system, depth can be estimated with below steps:

·???????? Camera Calibration

·???????? Disparity Calculation

·???????? Depth Estimation

But due to their indirect way of depth estimation, these systems can provide depth with good accuracy when calibration and disparity calculation are precise.

In Stereo system, distance of object from camera can be calculated using below formula:

???

For 10mm focal length, 20mm baseline and 5 micron/pixel, stereo system can measure a maximum distance of 40 meter.

As shown in above graph, because of the inverse relationship between disparity and depth (distance), stereo systems are reliable for a specific disparity range. Beyond this defined disparity range depth estimation may fail abruptly. Green highlighted area can be considered safer range to estimate depth. While red highlighted area which represent distant points shows larger variations in depth for every pixel change in disparity.

With mentioned configurations, disparity values below 8 pixels for disparity error of 1 pixel will lead to 50cm to 7 meter depth error. Disparity error depends on the accuracy of the matching algorithms which matches the points in the two images. Some algorithms can even achieve accuracy of sub-pixels while others of 1 pixel only.

Sub pixel calculations improve accuracy, still depth estimation of distant points accurately and reliably is difficult which makes stereovision system less reliable for ADAS.

It is true that Every Pixel Matters for vision based system as every life counts !!

RADAR based collision avoidance systems on the other hand can work in harsh weather conditions like pitch dark nights, in fog, in heavy rain etc. and have an excellent resolution in terms of angle , range and velocity .

Starkenn Technologies , Ashish Deshpande , Shashank Joshi , Ranjeet Bhelke , Sumedh Badve , Swastid Badve