AICity2020-Anomaly-Detection

Multi-Granularity Tracking with Modularlized Components for Unsupervised Vehicles Anomaly Detection (CVPRW 2020)

This repository contains source codes of team113 for NVIDIA AICity Challenge 2020 Track 4, and the technical details please refer to the paper "Multi-Granularity Tracking with Modularlized Components for Unsupervised Vehicles Anomaly Detection"

Our method obtain the F1-score metric at 0.9855 and the RMSE metric at 4.8737, which ranked first in the Track4 test set of the NVIDIA AI CITY 2020 CHALLENGE.

Requirements

• Paddle1.7-gpu-post97

• cuda9

• cudnn7.5

Annotations

The training annotations for detection models are in here. Extract one frame every 2 seconds, then choose half of them to annotate. The corresponding is (annotation_number - 1)*2*30 = frame_number

Step1: Detection Model

Train the detection model

cd det_code/PaddleDetection
sh train.sh

Inference Procedure

cd det_code/PaddleDetection
sh infer.sh

Step2: Background Modeling

Extract background

python bg_code/ex_bg_mog.py

Step3: Extraction of Hypothetical Abnormal Mask

Obtain motion-based mask

python mask_code/mask_frame_diff.py

Obtain trajectory-based mask

python mask_code/mask_track.py

Mask Fusion

python mask_code/mask_fuse.py

Step4: Box-level Tracking

Tube construction

python box_track/tube_construction.py

Box_level Tracking

python box_track/box_level_tracking.py

Step5: Pixel-level Tracking

Coarse anomaly result for Pixel-level Tracking

python pixel_track/coarse_ddet/pixel-level_tracking.py

Similarity filtering for the preliminary abnormal candidate results

python pixel_track/post_process/similar.py

Backtrack the start time

python pixel_track/post_process/time_back.py

Merge in the temporal dimension

python pixel_track/post_process/id.py

Step6: Fusion and Backtracking Optimization

Fusion of box_level tracking and pixel-level tracking, and backtracking

python fusion_code/fusion_backtracking.py

If you have any questions or issues in using this code, please feel free to

contact us ([email protected] or [email protected])