Storyboard

Artificial perception systems implement physical principles analogous to biological systems, but with different technologies and operating ranges. Digital cameras are the analogue of eyes: the fundamental noise is shot noise (shot noise) SNR = N_e, which states that for SNR = 100 10,000 photoelectrons are needed per pixel. The best scientific sensors have QE > 95% and N_e,min = 1 electron.

LiDAR (Light Detection and Ranging) is the artificial analogue of bat echolocation: it emits laser pulses and measures the return flight time. The angular resolution = /D is diffraction limited (same as the eye), but the distance resolution d = c·/2 can be millimeter with nanosecond pulses. LiDAR is the reference perception technology in autonomous vehicles.

The radar equation shows that the maximum range R_max (P_t·G²·_RCS)^(1/4): to double the range you need to increase the power 16 times. High-resolution radars (SAR, Synthetic Aperture Radar) exploit the movement of the platform to synthesize an aperture much larger than the physical antenna, achieving azimuthal resolution _az = D_antenna/2, independent of distance.

The IMU+GPS fusion using an extended Kalman filter is the standard for autonomous navigation: the IMU provides high-frequency information but with drift (analogous to the semicircular canals), while the GPS provides absolute position but at low frequency and with noise (analogous to visual landmarks). The fusion produces _pos < 1 m under normal conditions, comparable to the navigation accuracy of migratory birds.

Convolutional neural networks (CNNs) are the artificial analogue of the V1V2V4IT cortical visual hierarchy. The 2D convolution implements the local (translation-invariant) receptive field response, analogous to the simple cells of V1. Successive layers build representations of increasing complexity: edges textures parts objects, in parallel with the ventral visual hierarchy. Modern CNNs surpass human accuracy in image classification.

>Model

ID:('ky', 597)

Description

ID:(0, 597)