The approach is fine for general imagery (all images from different cameras ensembled) but to work in the car it must take sensor noise/gain profile into account. Dct will show quite high amplitudes at high frequencies at night (and during sunset scenes and even the image is blurred or lens is dirty).
When I approached same problem I designed band pass to filter to look for certain size features and sensor noise profile guided low-pass filter and compared the results of two outputs to harvest for dirty images (having only noise and no details) then trained a dnn to deploy