Single-camera and Inter-camera Vehicle Tracking and 3D Speed Estimation Based on Fusion of Visual and Semantic Features (Winner of Track 1 and Track 3 at the AI City Challenge Workshop in CVPR 2018)
This repository contains our source code of Track 1 and Track 3 at the AI City Challenge Workshop in CVPR 2018. Our team won in both of the tracks at the challenge.
The source code of Track 1 is built in MATLAB and C++, with our trained YOLOv2 model provided.
The source code of Track 3 is developed in Python and C++.
The code has been tested on Linux and Windows. Dependencies include CUDA, cuDNN and OpenCV.
The team members include Zheng (Thomas) Tang, Gaoang Wang, Hao (Alex) Xiao, Aotian Zheng.
[Paper], [Slides], [Poster], [Project Page], [2018 NVIDIA AI City Challenge]
The datasets for the 2018 AI City Challenge are no longer available to the public. However, as the 2019 AI City Challenge Workshop was launched at CVPR 2019, they provided a new city-scale dataset for multi-camera vehicle tracking as well as image-based re-identification. They also had a new dataset for traffic anomaly detection. The scale of the datasets and the number of vehicles that are being used for evaluation are both unprecedented.
To access the new datasets, each participant will need to download a data release form at the AI City Challenge website. Please forward your inquiry to aicitychallenge2019@gmail.com.
The NVIDIA AI City Challenge Workshop at CVPR 2018 specifically focused on ITS problems such as
- Estimating traffic flow characteristics, such as speed
- Leveraging unsupervised approaches to detect anomalies caused by crashes, stalled vehicles, etc.
- Multi-camera tracking, and object re-identification in urban environments
Our team participated in 2 out of 3 tracks:
- Track 1 (Traffic Flow Analysis) - Participating teams submit results for individual vehicle speed for a test set containing 27 1-minute videos. Performance is evaluated based on ground truth generated by a fleet of control vehicles that were driven during the recording. Evaluation for Challenge Track 1 is based on detection rate of the control vehicles and the root mean square error of the predicted control vehicle speeds.
- Track 3 (Multi-camera Vehicle Detection and Reidentification) - Participating teams identify all vehicles that are seen passing at least once at all of 4 different locations in a set of 15 videos. Evaluation for Challenge Track 3 is based on detection accuracy and localization sensitivity for a set of ground-truth vehicles that were driven through all camera locations at least once.
Detailed information of this challenge can be found here.
Our team achieves rank #1 in both Track 1 and Track 3. The demo videos can be viewed here.
In SCT, the loss function in our data association algorithm consists of motion, temporal and appearance attributes. Especially, a histogram-based adaptive appearance model is designed to encode long-term appearance change. The change of loss is incorporated with a bottom-up clustering strategy for the association of tracklets. Robust 2D-to-3D projection is achieved with EDA optimization applied to camera calibration for speed estimation.
The proposed appearance model together with DCNN features, license plates, detected car types and traveling time information are combined for the computation of cost function in ICT.
Under the Track1 folder, there are 6 components:
- 1_VDO2IMG: Converting each video file to a folder of frame images
- 2_CAM_CAL: Manual camera calibration based on minimization of reprojection error by EDA optimization
With the access to GPS coordinates (using Google Maps or other tools), you can use our newly developed PnP-based calibration tool here instead. - 3_YOLO_VEH: Extension of the YOLOv2 object detector with our trained model for vehicle detection/classification
We strongly encourage users to try the latest YOLOv3 object detector instead. - 4_TC_tracker: Proposed tracklet-clustering-based tracking method
Note that this SCT method has been upgraded into TrackletNet Tracker (TNT). The corresponding paper on arXiv is here. The source code (training + testing) is provided here. - 5_APP_MDL (optional): Extraction of histogram-based adaptive apperance models and their comparison
- 6_SPD_EST: Speed estimation based on input of tracking results and camera parameters
Detailed description of each package is given in each subfolder.
Under the Track3 folder, there are 3 components:
- 1_Multi-Camera Vehicle Tracking and Re-identification: Multi-camera vehicle tracking based on a fusion of histogram-based adaptive appearance models, DCNN features, detected car types and traveling time information
- 2_YOLO_LP: Detection of license plate from each cropped vehicle image based on YOLOv2 We strongly encourage users to try the latest YOLOv3 object detector instead.
- 3_LP_COMP: Comparison of license plates under low resolution
Detailed description of each package is given in each subfolder.
The output of 1_Multi-Camera Vehicle Tracking and Re-identification is the similarity score between each pair of vehicles for comparison. It can be converted into a distance score by inverse proportion. The output of 3_LP_COMP is the distance score between each two license plates. The final distance score between two vehicles is the multiplication of the above two distance scores.
Please consider to cite these papers in your publications if it helps your research:
@inproceedings{Tang18AIC,
author = {Zheng Tang and Gaoang Wang and Hao Xiao and Aotian Zheng and Jenq-Neng Hwang},
title = {Single-camera and inter-camera vehicle tracking and {3D} speed estimation based on fusion of visual and semantic features},
booktitle = {Proc. CVPR Workshops},
pages = {108--115},
year = {2018}
}
@misc{Tang17AIC,
author = {Zheng Tang and Gaoang Wang and Tao Liu and Young-Gun Lee and Adwin Jahn and Xu Liu and Xiaodong He and Jenq-Neng Hwang},
title = {Multiple-kernel based vehicle tracking using {3D} deformable model and camera self-calibration},
howpublished = {arXiv:1708.06831},
year = {2017}
}
For any question you can contact Zheng (Thomas) Tang.