I'm working towards getting into a Computer Vision-based tech team in college. To get in, I have to impress my college seniors with an independent project. The team works on autonomous driving, so any ideas for a project related to this field would help a lot.

I've completed the Andrew Ng course on Machine Learning on Coursera and also CS231n by Stanford University, so y'all can judge my skills based on that. This would be my first project under Computer Vision, but I need it to be good enough to impress my seniors. The problem I'm facing is that the project ideas on the internet already have been implemented using specific algorithms. I want to try and implement something original, but I'm not sure if I have the skills to do so.

Basically, I'm really confused, any help would be appreciated.

  • $\begingroup$ Impressing the college seniors in a computer vision course has to do with story telling. A character has to reach a goal node in the plot and avoiding other nodes. Formalizing this task is possible with a semantic network, in which the possible outcome is formalized. For example “if project=new then college senior=positive”. A first step is to search for a dataset, in which previous attempts of computer science students to succeed in academia are analyzed and then the raw data is converted into a model. This allows to anticipate future situations. $\endgroup$ – Manuel Rodriguez Nov 18 '19 at 7:58

Step one, engage with the seniors.

Since, self-driving is a complex task the seniors are (with high probability) using a pipelined system with multiple modules. Ask them for a ceiling analysis. If they do not know what that is (you've already wowed them). Assuming that they know what ceiling analysis is, they can provide you with one for one of their projects. Use this analysis to figure what would be the best part of their pipeline on which to focus. I also recommend asking them for the learning curves of these systems (if they have them) if not that's something you could work on.

Additionally, vision is a complex task even in isolation. But here are some ideas that will hopefully provide a starting point:

The driving task is noisy, the light might be green but there may be a pedestrian illegally crossing the street. The system should always favor human safety above all else. Another example might be to spot if another driver is not focusing on the road - maybe through a side view camera you can look for people that are texting and driving and hence alert the vehicle to avoid driving next to the car.

Additionally, if a driver has demonstrated unsafe behavior the system should temporarily remember this car/driver and avoid them - it could store an encoding of the car by slowing down and grabbing the make, model & license plate (there are lots of blogs/articles about one-shot learning that could be applied here).

Another idea, perhaps a drunk driver is on the wrong side of the road - the system should be able to recognize "unsafe situations" and perhaps have a classification task of the "best action." This would be akin to a knee-jerk reaction - that is, the car's vision system should have rapid reflex actions that require speedy processing. For example, a dear jumps out unto the road the system should output the reflex: [brake]. There are many edge cases: The system should observe if the driver is wearing a seatbelt before it slams on the brakes - this might imply that it would be safer sometimes to engage the airbag first then brake.

I would recommend brainstorming for about 20-30 minutes to see if you can think of edge cases and how you might solve these with a vision system.

I hope this helps, and best of luck!

| improve this answer | |

Not the answer you're looking for? Browse other questions tagged or ask your own question.