Driver Behavior Cloning using Deep Learning for Level 4 Self-driving Vehicles

There are various scenarios on road with different curvatures, turns, slopes, bridges, flyovers and others along with different obstructions. A human driver analyzes these situations while maneuvering or driving the vehicle.
An algorithm is built to teach a self driving vehicle to drive on its own by cloning the actions a human driver executes while steering the car in various scenarios.

A CNN (Convolutional Neural Network) algorithm maps the actual data from the steering wheel while the driver steers the vehicle, with the images of the road curvature captured from the cameras.

This CNN Network Algorithm built using Deep Learning model learns on its own to build a stronger algorithm to achieve the required precision and accuracy for the self-driving vehicle. The algorithm clones the human behavior to drive the vehicle on its own.

Our approach

Phase 1 – Training phase

The actual steering data is compared with the road curvature images captured from three cameras mounted on the car to train the CNN Network

The CNN Network maps the curvature of the road with the corresponding steering angle (actual rotation and shift angle of steering wheel)

Phase 2 – Deployment phase

The CNN Network built during training phase is deployed in the self-driving vehicle to generate the steering rotation angle using the road curvature data captured only from the center camera

The CNN Network can generate the necessary commands to the steering wheel to adjust the rotation & steering angle as per the road conditions (curvatures)