So for the "at night" part of this, I have some thoughts -
1) Brandon is right that the visual indicators are more effective at night, both passive (reflectors) and active (lights).
2) The CG must contain both passive and active visual indications... that is, reflectors and lights.
3) The vision system can be made to work at night, but it will have challenges compared to day time. The biggest being that the dynamic range looking backwards is huge and can washout the sensors - the very bright headlights on a background of, at worst case, total blackness. More confusingly, a procession of cars, as seen by the CG may just look light a row of any number of lights. Keep in mind the cases where a vehicle is missing a headlight, or a motorcycle is approaching from the rear. Keep also in mind the case when two vehicles are approaching, but the furthest is partially obscured by the closest. Solutions here include modifying the CNNs to train on varying light levels and headlight configs, but also to work with the sensors, image processing, and infra-red capabilities. There may also be some trickery that could be done with masking headlights and adjusting dynamic range after that. There are solutions from problems reading reflective license plates at night which have addressed problems like this.