IMPORTANT
This project is just a template. It does not contain any working code. The development of this project based on the mentioned paper was stopped due to changes of priorities in our team. We apologize for the inconvenience.
This repository contains a customized source code of Track 4 in the NVIDIA AI City Challenge at CVPR 2021.
The Track 4 of NVIDIA AI CITY 2021 comprised of detecting anomalies based on different video feeds available from multiple cameras at intersections and along highways.
The paper of this code can be found in arxiv
Detailed information of NVIDIA AICity Challenge 2021 can be found here.
Overview of the architecture of our anomaly detection framework, which consists of three main pipelines.
- Python 3.7
- To run the model from scratch, please request data from the organizers.
- PaddleDetection
- Run
python install_decord.py
.
To run model from scratch, please follow these steps:
-- Pre Processing
- Run
python pre_processing/stabilize_video.py
to apply DVS. - Run
python pre_processing/model_background.py
to segment the processed frames. - Run
python pre_processing/detect_vehicle.py
for two-stream vehicle detection. - Run
python pre_processing/generate_mask.py
to produce trajectory-based masks.
-- Dynamic Track
- Run
python dynamic_track/pixel-level_tracking.py
to filter out suspicious events. - Run
python dynamic_track/intra-tube_judgement.py
to remove the parts that do not belong to current anomalous vehicles. - Run
python dynamic_track/inter-tube_fusion.py
to fuse candidate tubes which represent to the same vehicle.
-- Post Processing
- Run
python post_processing/detect_vehicle_collision.py
to detect vehicle collisions. - Run
python post_processing/refine_temporal_boundary.py
to obtain a more accurate temporal localization.