Self-Driving Car Engineer Diary — 12
Sat, 16/Sep/2017
Term 3, month 2 and discovered that we were allowed to do both electives, so I jumped straight in to see if I could complete them within the time box.
Advanced Deep Learning Elective
Fully Convolutional Networks (FCNs) output an entire image classified at the pixel level (‘semantic segmentation’) instead of simple object detection / classification (like with YOLO). This granularity identifies free driving space around the vehicle (feeding into path planning) and helps produce enhanced object maps. Combining Depth and Segmentation decoders may be a feasible replacement of LIDARs. Deep Learning is an amazing area of AI!
To apply the FCN architecture above, we went through the following build sequence :
1. Load the pre-trained VGG16 object model and access its layers (1–3, 4, 7).
2. Encode output layers (3, 4, 7) to 1x1 convolutional layers.
3. Combine skip-layer 4 to existing output layer 7, then take that combined output and add skip-layer 3 during upscaling (aka Decoder) process.
4. Final upscaling operation to get original image size.
5. Optimised the network classifying [road | not road] pixels, batch trained the network while playing with the hyperparameters.
Functional Safety Elective
This elective was a real surprise! I was expecting that application of ISO 26262 standard would be a little dry but when viewed with the thought of …
“how can I architect an autonomous system to reduce injury / fatality risk to my loved ones?”
… the course content suddenly became very real and compelling.
Focusing on currently available Lane Assistance safety features like Lane Departure Warning (LDW) and Lane Keeping Assistance (LKA) the elective taught us how to plan, hazard scenario brain-storm / risk assess and solution architect / define our way to ISO 26262 compliance.