Skip to content

This project aims to recognize and track escooters through the traffic intersection and find their trajectories, time of passing and speed.

Notifications You must be signed in to change notification settings

savenow/Traffic_Camera_Tracking

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Traffic Camera Tracking (SAVeNoW)

Basic Overview


Main goals of this project is to detect, track and obtain velocity information of seven different traffic actors (Escooter, Pedestrians, Cyclists, Motorcycle, Car, Truck, Bus) from a video footage of a traffic intersection in Ingolstadt. This project is a part of SaveNoW, where all these data are used to create a digital twin of Ingolstadt for simulation purposes.

Individual Components


Object Detection:

(Still actively developed..A lot of the parameters could change in future)

PyTorch based YoloV5 framework is used for detection and the model yolov5l6 is trained on custom dataset of around 15k images (13k train/1.5k valid) in varied conditions (daylight, rain and night) using custom data, cityscapes and traffic data from COCO. All the images were manually labelled using CVAT online annotation tool. The model was trained by for 375 epochs using RTX 3090 at img size 1920 for multiple days to achieve final accuracy of 77%

For faster inference speeds, a TensorRT Engine has been built from the traied weights and can infer at 125 fps using RTX 3090

Tracker:

Since YoloV5 is spatial CNN Object detector, it doesn't contain temporal information to recognize an object across multiple frames. For this purpose, a modified SORT (Simple Online Realtime Tracker) has been used.

Postprocessing:

The inference+tracker data is inherently noisy. So following postprocessing steps are applied to make it smooth and consistent:

  1. Removing temporary noisy trackers
  2. Interpolating missing tracker position
  3. Estimating velocity for each tracker based on camera-image homography data and smoothening the trajectory with rolling average
  4. ClassID matching to predict which class the tracker most probably belongs to. This also prevents classid switching of trackers
  5. Visualize heading angle
  6. Converting image coordinates to lat-long coordinates (again based on camera-image homography)
  7. Finally visualizing it in a video

Sample inference:


Usage:

python inference.py --input {INPUT_VIDEO_PATH} --model_weights {WEIGHTS_PATH} --output {OUTPUT_PATH} --minimap --trj_mode

Notes


  • Weights files and training datasets are not present in the repo. If you are interested, please feel free to contact us: pascal.brunner@carissma.eu
  • Calibration values must be properly tuned depending on the individual scene (Hint: Maybe will be automated in the future)

Contributing


Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

About

This project aims to recognize and track escooters through the traffic intersection and find their trajectories, time of passing and speed.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages