🥉News: our team got the 3rd place in the AICity 2021 Challenge on Track 4

Dual-Modality Vehicle Anomaly Detection via Bilateral Trajectory Tracing (CVPRW 2021)

Paper Link

Paper-with-code

ArXiv Link to the Paper

Introduction

This is the source code for Team WHU_IIP for track 4 Anomaly Detection in AICity 2021 Challenge.

Our experiments conducted on the Track 4 testset yielded a result of 0.9302 F1-Score, and 3.4039 root mean square error (RMSE), which performed 3rd place in the challenge.

Fig1 Rank of our team

More implementation details are displayed in the paper—— Dual-Modality Vehicle Anomaly Detection via Bilateral Trajectory Tracing

The paper link will be added after CVPRW2021. Here we only show the flow chart for a better understanding of the following procedures.

Fig2 Flow Chart

Our Solution for NVIDIA AICity Challenge 2021 Track4

Requirements

• Linux (tested on Ubuntu 16.04.5)
• Packages (listed in the requirements.txt)

Annotations

We have annotated 3657 images selected from training dataset. The training set and testing set are randomly split at the ratio of 4:1. The link for annotation is included as follows.

Annotations link: Google drive

Procedures

Background Modeling

Extract the background

cd bg_code
python ex_bg_mog.py

Preparation For Detection

Structure of PreData Folder

The original videos and their frames are put under ../PreData/Origin-Test and ../PreData/Origin-Frame folders, respectively. And the background modeling results are placed under the ../PreData/Forward-Bg-Frame folder.

All these files are organized for the Detect Step later. Then the detection results based on background modeling will be saved under ../PreData/Bg-Detect-Result/Forward_full for each video while ../PreData/Bg-Detect-Result/Forward is kept in frames separated from full videos.

The detailed structure is shown below.

├── Bg-Detect-Result
│   ├── Forward
│       └── 1
│          ├──test_1_00000.jpg.npy
│          ├──test_1_00001.jpg.npy
│          ├──test_1_00002.jpg.npy
│          └── ...
│       ├── 2
│       ├── 3
│       └── ...
│   └── Forward_full
│       ├── 1.npy
│       ├── 2.npy
│       ├── 3.npy
│       └── ...
├── Forward-Bg-Frame
│   ├── 1.mp4
│   ├── 2.mp4
│   ├── 3.mp4
│   └── ...
├── Origin-Frame
│   └── 1
│       ├──1_00001.jpg
│       ├──1_00002.jpg
│       ├──1_00003.jpg
│       └── ...
│   ├── 2
│   ├── 3
│   └── ...
├── Origin-Test
│   ├── 1.mp4
│   ├── 2.mp4
│   ├── 3.mp4
│   └── ...

Detection

For detection model training instruction, please view the official yolov5 repo

Road Mask Construction

Extract Motion-Based Mask

cd mask_code
python mask_frame_diff.py start_num end_num

Extract Trajectory-Based Mask

python mask_track.py video_num

Mask Fusion

python mask_fuse.py video_num

Pixel-Level Tracking

Coarse Detect

cd pixel_track/coarse_ddet
python pixel-level_tracking.py start_num end_num

Fuse Similar Results

cd pixel_track/post_process
python similar.py start_num end_num

Filter Suspicious Anomaly Results

python filter.py

Fuse Close Results

python pixel_fuse.py

ROI Backtracking for Pixel-Level

python timeback_pixel.py type_num start_num end_num

Fuse Tracking Results

python sync.py

Box-Level Tracking

ROI Backtracking for Box-Level

Dynamic Analysis Stage

We mainly contribute this to trace the exact time of crashing since what's done before can only be used to locate the time when abnormal vehicles become static.

Multiple Vehicle Trajectory Tracing

cd car_crash
python crash_track.py

Singular Vehicle Trajectory Tracing

Demo

Multiple Vehicle Trajectory Tracing

Statistically, vehicle crashes often come up with sharp turns, which is the primary reaction of drivers when encountering such anomalies. Here we list some typical scenarios to display that.

Singular Vehicle Trajectory Tracing

Citation

@InProceedings{Chen_2021_CVPR,
    author    = {Chen, Jingyuan and Ding, Guanchen and Yang, Yuchen and Han, Wenwei and Xu, Kangmin and Gao, Tianyi and Zhang, Zhe and Ouyang, Wanping and Cai, Hao and Chen, Zhenzhong},
    title     = {Dual-Modality Vehicle Anomaly Detection via Bilateral Trajectory Tracing},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {4016-4025}
}

If you have any question, please feel free to contact us. (jchen157@u.rochester.edu and yuchen_yang@whu.edu.cn)