Pedestrian intention classifier is a human pose based classifier which achieves 80+% accuracy on JAAD dataset. The purpose of this classifier is to explore the machine learning on pedestrian crossing road intention for autonomous vehicle.
80+% accuracy on JAAD dataset.
Some false positive observed due to pose feature inherent lack of context infomation.
- Train image sequences (14 frames) as input instead of single image
- Add context info as feature such as road segmentation and traffic light.
MIT