A Neurocalibration Model for Autonomous Vehicle Navigation

The paper evaluates the capability of a neural model to calibrate a digital camera. By calibrate we understand the algorithms that reconstructs the 3D structure of an scene from its corresponding 2D projections in the image plane. The most used 3-D to 2-D geometrical projection models are based in the pin-hole model, a free distortions model. It is based in the correspondence established between the image and the real-world points in function of the parameters obtained from examples of correlation between image pixels and real world pixels. Depending on the sensor used, different kind of chromatic aberrations would appear in the digital image, affecting the brightness or the geometry. To be able to correct these distortions, several theoretical developments based on pin-hole models have been created. The paper proves the validity of applying a neural model to correct the camera aberrations, being unnecessary to calculate any parameters, or any modelling. The calibration of autonomous vehicle navigation system will be used to prove the validity of our model.