VTPLSTM
所属分类：自动驾驶
开发工具：Python
文件大小：0KB
下载次数：0
上传日期：2021-01-04 13:19:12
上 传 者：
sh-1993
说明： 灵感来自Social LSTM的表现。我创建了一个模型，可以在5秒内预测车辆的轨迹。该模型以LSTM为中心...，
(Inspired by the performence of Social LSTM. I created a model that could predict vehicle s trajectory in 5s. The model uses LSTM as center. I trained it at NGSIM. Still fix it...)
文件列表:
data(sample)/ (0, 2021-01-04)
data(sample)/test.csv (364275, 2021-01-04)
data(sample)/train.csv (364275, 2021-01-04)
data_generator.py (5424, 2021-01-04)
data_loader.py (11344, 2021-01-04)
log/ (0, 2021-01-04)
log/2020-10-23-15-47/ (0, 2021-01-04)
log/2020-10-23-15-47/model.txt (18657, 2021-01-04)
log/2020-10-23-15-47/net.pkl (834164, 2021-01-04)
log/2020-10-23-15-47/test/ (0, 2021-01-04)
log/2020-10-23-15-47/test/vdlrecords.1603439249.log (75942, 2021-01-04)
log/2020-10-23-15-47/train/ (0, 2021-01-04)
log/2020-10-23-15-47/train/vdlrecords.1603439249.log (75942, 2021-01-04)
log/2020-10-24-15-10/ (0, 2021-01-04)
log/2020-10-24-15-10/model.txt (37295, 2021-01-04)
log/2020-10-24-15-10/net.pkl (1034866, 2021-01-04)
log/2020-10-24-15-10/test/ (0, 2021-01-04)
log/2020-10-24-15-10/test/vdlrecords.1603523423.log (155945, 2021-01-04)
log/2020-10-24-15-10/train/ (0, 2021-01-04)
log/2020-10-24-15-10/train/vdlrecords.1603523423.log (155945, 2021-01-04)
log/2020-12-28-09-32/ (0, 2021-01-04)
log/2020-12-28-09-32/model.txt (55456, 2021-01-04)
log/2020-12-28-09-32/net.pkl (97222, 2021-01-04)
log/2020-12-28-09-32/test/ (0, 2021-01-04)
log/2020-12-28-09-32/test/vdlrecords.1609119171.log (235545, 2021-01-04)
log/2020-12-28-09-32/train/ (0, 2021-01-04)
log/2020-12-28-09-32/train/vdlrecords.1609119171.log (235545, 2021-01-04)
model.py (9288, 2021-01-04)
parameters.py (2862, 2021-01-04)
train.py (8171, 2021-01-04)
utils.py (7999, 2021-01-04)
visualize.py (3375, 2021-01-04)
图片2.png (47465, 2021-01-04)
图片3.png (56225, 2021-01-04)
**Project state: Still Fixing**
Inspired by the performence of Social LSTM. We created a model that could predict vehicle's trajectory in 5s. The model uses LSTM as center. We are still working on improving its performance!
**Implement detail:**
Baseline implementation: [https://github.com/quancore/social-lstm](https://github.com/quancore/social-lstm)
Baseline paper: [http://cvgl.stanford.edu/papers/CVPR16_Social_LSTM.pdf](http://cvgl.stanford.edu/papers/CVPR16_Social_LSTM.pdf)
**Documents Introduction**
**Data(sample):** There are two .csv files obtained by data_generator.py. The number of sequences is too small for study, so you need to generate dataset in your computers.
**Log:** Including our training result and the curve of loss.
**Parameters.py:** All of the setting are in this files including train, long-term train, visualize, data generator.
**Data_generator.py:** To generate sequences for training and visualize.
**Data_loader.py:** To read .csv files and execute them. (Normalization, social pooling…)
**Model.py:** Kernel file, including the definition of our model
**Train.py:** The entrance of training.
**Visualize.py:** Using matplotlib to visualize predicted trajectory.
**Model information**
VPTLSTM(
(cell): LSTMCell(64, 32)
(input_embedding_layer): Linear(in_features=9, out_features=32, bias=True)
(social_tensor_conv1): Conv2d(32, 16, kernel_size=(5, 3), stride=(2, 1))
(social_tensor_conv2): Conv2d(16, 8, kernel_size=(5, 3), stride=(1, 1))
(social_tensor_embed): Linear(in_features=32, out_features=32, bias=True)
(output_layer): Linear(in_features=32, out_features=5, bias=True)
(relu): ReLU()
(dropout): Dropout(p=0, inplace=False)
)
**Visualization**
The predicted trajectory is correct but it lack of accuracy on intention recognition. Maybe there are too less vehicles to change lane so that the model couldn’t get the conditions of changing lane. We are trying to get more relevant datasets about it.
(The solid line is the real track and the dashed line is the predicted track)
**Requirements:**
Pytorch
Numpy
Matplotlib
Pandas
visualdl
近期下载者：
相关文件：
收藏者：