The input values are image arrays 120x160 pixels with 3 channels for red,green,blue. The values range from 0-255 and are not normalized before they are fed into the convolution layer. I found this did not make a difference.

The output of the model is a single real number between -90(left) and 90(right). I believe a better approach would be to bin the outputs and use a classifier. This way you'd know when the model was getting confused (ie, approaching a perpendicular line.