Detects Chinese traffic police commanding poses 检测中国交警指挥手势
ChineseTrafficPolicePose is a network that classify
8 kinds of Chinese traffic police commanding poses by analyzing visual information.
ChineseTrafficPolicePose 是一个仅依靠视觉信息区分8种中国交警指挥手势的网络
This code runs under tensorflow 1.4, it’s hard to build now because Tensorflow has changed it’s API a lot. A pytorch version of police gesture recognizer is being maintained with pretrained models available at:
基于pytorch的、有预训练模型的版本:
https://doi.org/10.1016/j.neucom.2019.07.103
We publish the Police Gesture Dataset, which contains the videos of Chinese traffic police commanding gestures, and ground truth gesture labels for each video frame.
Police Gesture Dataset Download link: Google Drive
Notice: This gif is outdated. current version support prediction for FULL BODY, include legs. Check the videos in our dataset for examples of supported videos.
Watch Videos:
Environment
Python3
Tensorflow
with GPU supportTraining
"train", "test_a", "test_b", "val"
.parameters.TRAIN_FOLDER
. You may change the content of this parameter according to your path.python3 PAF_train.py
to train the keypoint network.dataset/csv_train
and dataset/csv_test
.dataset/policepose_video
.python3 PAF_detect.py dataset/policepose_video -a
to parse videos to skeletal data.python3 rnn_train.py
to train LSTM using labels from dataset/csv_train
and skeletal data from ./dataset/gen/rnn_saved_joints
.python3 rnn_detect.py -p
to predict test videos using name list from dataset/csv_test
and skeletal data from ./dataset/gen/rnn_saved_joints
.Python3 rnn_detect.py -e
to print Edit Distance of predicted labels with ground truth labels from dataset/csv_test
.