Issues Log

Log No. Date Discovered Date Resolved Description Issue Resolution Update/Status
02/11/202004/11/2020Data capture- 3D to 2D homography outputs of bounding boxes are not good.Problem in world to camera lens transformation was incorrect. Fixed.
03/10/202003/20/2020Constant acceleration motion model causing problemsMotion model changed to constant velocity in pixel space
03/17/202004/03/2020Problems with tracking association.Switched to scipy's implementation of tracking association.
4. 04/18/202008/31/2020Simulation speed at the edge of requirements.Using Carla's traffic manager for the core project goals and using batch inference for the learning based model inference.
509/15/202009/30/2020Learning Based model might not be enough for achieving the project requirementsFocus on the metric extraction and rule based modelling using the newer Traffic Manager
610/20/202010/25/2020Newer version of Carla has a different map numbering system and also has updated camera parameters which cause troubles in the data capture pipelineWe decided to re capture the traffic light ids and also manually recompute the correspondences between the camera and the bird's eye view image
710/15/202010/22/2020There were certain issues with the lead vehicle matching algorithm where we were matching vehicles in other lanes due to noisy velocity estimates, similar issues were observed in the tracking system as wellWe concluded that having a bird's eye view HD map with lane positions would be helpful and reduce noise. This solution worked really well.
810/27/202011/12/2020Using the bottom center of the 2d bounding dox for transforming the vehicle coordinates causes some positioning edge casesWe are working on using 3D bounding box predictions and use the actual "base" of the vehicles enclosing cuboid to transform the vehicle
911/05/202011/20/2020Our learning based modelling performance is limited by Carla.This can be resolved as it would require changes to Carla code.