The final product is here, http://asherman.site/hazcam/ and I also release the trained haar classifier separately on my own GitHub in case anyone had another useful projects for it. It was a lot of work hand selecting and cropping 400+ photos of the rear ends of cars, so I figured I might as well share the result. https://github.com/pddenhar/OpenCV-Dashcam-Car-Detection
I was recently hit by a Mercedes SLK, and it's not something I'm anxious to repeat.
At the moment I'm thinking a GoPro streaming wirelessly to a Raspberry Pi, which can do the visual processing and essentially warn me if something big and metal is coming up behind me too quickly. I got the hardware, but I really don't have the software chops so if someone could point me in the right direction I would really appreciate it!
There are a lot of tutorials online on developing computer vision with the PI.
However, it sounds like the K.I.S.S. solution for you would be a backup camera or something so you can see whats happening behind you instead of relying on CV (computer vision) especially if you don't have the software chops to take on the CV challenge.
The Raspberry Pi is going to be the ideal machine I think to do the heavy lifting hopefully, and the main reason I am open to rely on CV is that there is nowhere to fit a screen on the front of my wheelchair. I tried streaming wirelessly from a go pro to my iPhone and the battery life was absolutely abysmal.
Maybe if the CV detected a problem it could make a fit bit or something similar vibrate to alert me to a problem, different numbers of vibrations were different situations etc etc. Although it is ENTIRELY possible I'm overthinking this!
The HOG + SVM method is quite slow and not as accurate as a deep learning approach. Before jumping onto semantic segmentation, I recommend re-implementing this project or more generally solve this problem using a Regional Convolution Neural Network architecture (R-CNN) like Faster R-CNN or YOLO for instance.
I wrote a realtime human detection library  for a robotics project that used HOG + a simple neural net for classification. While it worked okay, I wasn't happy with the precision (around 90%) and decided to try out a simple convnet from Torch (doing the classication on depth images instead of HOG descriptors). The Torch version was slightly slower on a CPU, but both the precision and recall jumped up drastically.
Second paragraph - 'argures' should be 'argues', and 'resonse' should be 'response'.