# [Read while you drive - multilingual text tracking on the road](https://link.springer.com/chapter/10.1007/978-3-031-06555-2_51)
The authors propose an end-to-end tracking model which utilises textual information.
They also benchmark the performances of several tracking and detection models.

### Proposed Method :

- The authors extend the RoadText-1K dataset to RoadText-3K dataset with 1000 images from spain and 1000 from India
- The proposed model uses Tracking by detection using CentreNet architecture.
- CentreNet is an OD model which represents objects as a single point in the bounding box center and regresses width and height at the cente of location.
- The centre is represented as Gaussian Kernel on a heatmap
- The model uses Mish activation instead of ReLU and adds one additional GRU cell before upsampling
### Experiments :
- The datasets used is RoadText-3k
- The metrics used is Multiple Object Tracking accuracy(MOTA), MOT precision(MOTP), ID switches and IDF1.
- The models evaulated are CTPN,EAST,FOTP and CRAFT
- The model is evaluated on frame level text detection and tracking.
- The proposed method achieves comparable results to the SOTA methods(FOTS and CRAFT)