The Belgium TS Dataset may be helpful, as well as The German Traffic Sign Detection Benchmark.
Additional Notes Based on Question Author's Idea
The idea in the question author's addendum of placing signs onto street sides and corners is a good one, but to do it repeatably and in a way that doesn't bias the training is its own research project. However, it is a good research direction. What would be of benefit to AV researchers worldwide is a multi-network topology and equilibrium strategy with the objective to create the following data generation features.
- Street sign symbol inputs in image form, with or without cropping, as movie frame sequences or single still shots, or from SVG files.
- Annotation generation using partially human-labelled data.
- 3D analysis of sign angle and perspective setting so that the images appear exactly as they would from a vehicle's imaging system.
- Matching of lighting between the superimposed sign and the background scene.
- Automatic blue-screening for the sign image.
This is obviously not a basic data hygiene problem. It is its own AI project, but the return on this research project in terms of furthering the AV technology is immense and may have significant data set statistical advantages over collecting data from the vendors that supply images to Google maps and other Big Data aggregators.