Hazard Detection for Self-Driving cars using Monocolar camera feed

Uses the Tensorflow Object Detection API to run the following models

• Object detection and classification according to predefined classes producing bounded boxes, object classes and detection scores.
• Multiple Object Tracking of Detected objects
• Trajectory prediction for each detected object
• A 3-Dimensional location (x,y,z) estimator using information from the object detection (bounded boxes) of all detected objects.

Requirements

• Python 3.6+
• Tensorflow 2.*
• PIP

Running the models

The main notebook to run is the hazard_detection notebook. This contains everything mentioned above in a single notebook. GPU is required for inferences to be made in good time.

Dataset

The dataset is configured for the KITTI Tensorflow Dataset, stored in a private Google Cloud Storage. (May be removed in a few months)

Demo

Acknowledgements

Uses SORT Algorithm by Alex Bewley

Uses some parts from KITTI distance estimation