Safe Goodbye

TEAM: Arrival(Jungyeon Lee/Chaeeun Kim/Yookyung Jeong)

If you click the image below, you can see the presentation video on Youtube.

Dataset

자율주행 - 버스 승객 승하차 영상 : Camera C (General C)

• Removed List

  >>> image problem
  ...\ele\[ele]attend_270C
  ...\etc\[etc]attend_154C
  ...\hospital\[hospital]attend_392C
  ...\hospital\[hospital]leave_251C
  >>> label problem
  ...\apt\[apt]attend_1275C
  ...\apt\[apt]attend_941C
  

• Target(demonstration result) from Validation set

reselecting

type 1
[hospital]attend_067C

type 2
[mid_high]leave_354C

	Train	Valid
per Scene
per Frame

Modeling

• 하차 전 약 35프레임에서의 각 사람의 skeleton data를 time series data로 만들어서 하차할 것인지 하차하지 않을 것인지 Intention Prediction

	Train	Valid
Loss
Accuracy

Codes

• data_analysis.py: 1
• make_dataset.py: 2
• dataset.py: 3
• simple_dataset_keras.py: 4 simple lstm model로 training
• train.py: 5
• generate_video.py: 6
• tf2_multipose.py: todo

Reference

• Scale invariant angle label
• seaborn.histplot
• matplotlib barplot
• AngleAnnotation class
• dealing with csv file
• Keras data generator
• Guide to the Functional API
• DataGenerator
• Tensorflow Keras - 4 (자연어처리,감정분석)
• Tensorborad
• Checkpoint
• ModelCheckpoint

Baseline

• Unified Framework for Pedestrian Detection & Intention Classification
• FuSSI-Net: Fusion of Spatio-temporal Skeletons for Intention Prediction Networks

OD+Tracking

• Yolov5 + Deep Sort with PyTorch
  • python track.py --source [village]day_001B.mp4 --yolo_weights yolov5/weights/crowdhuman_yolov5m.pt --classes 0 --save-txt --save-vid
  • frame_idx, id, bbox_left, bbox_top, bbox_w, bbox_h, -1, -1, -1, -1)
• YOLOv4-Cloud-Tutorial

Pose Estimation

• Deep High-Resolution Representation Learning for Human Pose Estimation (CVPR 2019)
• Multi Person PoseEstimation By PyTorch
  • python Demo_video.py -backbone {CMU or Mobilenet} -video {video path} -scale {scale to image} -show {}