[TLD] Train DNN for traffic light detection

[agotterba]

Train the traffic light detection network.

We can get data from the simulator, the test lot's bag file, or from these sources:

Bosch dataset: https://hci.iwr.uni-heidelberg.de/node/613278, https://github.com/bosch-ros-pkg/bstld68 Viva: http://cvrr.ucsd.edu/vivachallenge/index.php/traffic-light/traffic-light-detection/52

This is split into two parts:

1. Simulator model
2. Parking lot (Carla) model

[agotterba]

I was able to run detection on the ros bag file, and am seeing a low rate of detection. Attached is a selection of images from one run, along with the log file:

parking_lot_data.tar.gz

The images are named:

tcimage__.jpg

I set the score threshold to 0.5 (I previously had it at 0.65, but that resulted in too few detections). A score of 0.000 means there was no detection. Some, such as tc_image_1_green_0.586.jpg are classified incorrectly (in this case, the light is red). In ic_image_32_green_0.675.jpg, the light is red, and the score is reasonably high.

One thing I'd like to try in tl_classifier.py is cropping out the reflections on the hood of the car. I'll try that this weekend.

[jodiealaine]

Trained models are included in the following approved/merged PR:

• TLD new parking lot models