Abnormality Detection with Multi-Adaptive Vehicle Detectors for Traffic Video Analysis

This is the source code for Track 3 Ai City Challenge 2019.

Extract Road Mask: Tools/mask_creating.py

Extract Unchanged Scenes: Scene-Change-Stop-Detection

Training Retina Net: Detectron

Preprocessing data: link

Testing

To reduce the testing time:

• We saved all unchanged scene intervals in json format.

• We saved the result of day/night detector and hard code in Detectors.py

• We saved all road masks in npy format.

• We saved all detected bounding boxes. Because we use multiple detectors to finish the task, each one will have specific format to store detected bounding boxes. The code to parse detectors output is in Detectors.py

There are 3 classes to parse detector output in Detectors.py: Multiple adaptive vehicle detectors by Tran, RetinaNet trained on night videos AICity 2019, FRCNN's Jia Yi Wei, FRCNN trained on dashboard camera datasets.

Link download our result: Uploading

To reproduce the result: Run Test.py: It will takes unchanged scenes, masks, detector result, and external configuration to detect anomalies.

The full result content:

• Images of detected bounding boxes.

• Images of anomaly events and anomaly proposals.

• Confident score for each frames in 2 representations: graph and text

• Detected anomaly events in text: <video_id> <scene_id> <start_time> <end_time> <confident_score>

To merge detected events: Run ResultRefinement.py

Training

Mulitple Adaptive FRCNN: The training code and model can be found on HCMUS repository

Retina Net: We use the Detectron libraries to train model. The modified code inside folder Detectron

FRCNN Trained on dashboard camera dataset: We are not allowed to publish this source code because it belongs to other projects (update when available).

FRCNN's Jia Yi Wei: 2018AICITY_MCPRL