SSD Traffic Lights Detection

update:

Now we have a more advanced implementation for traffic light detection, the source codes can be found at strangeai.pro: link

it's more accurate, more faster, and more robustic in traffic light detection, also need fewer memory usage.

Prepare Data

This step is the same as my previous repo kitti_ssd, train kitti dataset on ssd, this repo using dataset is LISA, a traffic light detection dataset. As the same of previous, we get Images and Label folder which contains all images and labels, one image reflect on same named label file, in ./data folder we have a script gen_all_labels.sh to generate same named label files contains nothing, just in case some images has no label bound box, this may cause unable to train SSD.

Generate lmdb data

To train caffe, we'd better use lmdb database to feed data into network, since we got images and labels, oh, another thing I just forgot, label file format as follow:

class_index x_min y_min x_max y_max

Continue, to generate lmdb data you only need to type:

bash ./data/create_list.sh

This will generate trainval.txt , test.txt and test_name_size.txt. In create_list.sh you have to change your dataset name, and we suppose your data placed in ~/data. So, if you got a dataset named FACE, then your trainval.txt and test.txt will seems like this:

Images/0001.jpg Labels/0001.txt

Simple, right? and then:

bash ./data/create_data.sh

In the sh file, we have a data_dir viraible, please make sure data_dir + trainval.txt line can reach your real image, otherwise you cannot generate lmdb file, or just get zero lenght. After this step, you are going have trainval.lmdb and test.lmdb under ~/data/YOURDATASET/lmdb

Predict Using SSD

ok, before train SSD, you can test this, you can download 2 model, VGG16.v2.caffemodel and VGG_LISA_SSD_414x125_iter_120000.caffemodel`, the first one is SSD pretrain model, you gonna need it when you train SSD, and second is I trained model using for traffic light detection. Place first model into ./models/VGGNet/LISA/SSD_414x125 Place second model into ./models/VGGNet And then type:

python ssd_detection.py

You will see something like this:

Train SSD

To train SSD, you just need:

python train_ssd_lisa.py

But you need to change some directory and datset names.

Predict Many Images

We provide a script to predict all images under an dir, using ssd_detection_all.py instead.

Copyright

(c) DiDiChuXing All Rights Reserved