BigSnarf blog

Infosec FTW

Predicting deep into the future with segmentation

Self driving car LIDAR and camera download

Screen Shot 2017-03-23 at 11.06.36 PM

Self Driving Car Operating System

Screen Shot 2017-03-20 at 6.20.31 PMScreen Shot 2017-03-20 at 6.18.10 PMScreen Shot 2017-03-20 at 6.03.27 PM


TACC Traffic-Aware-Cruise-Control


Screen Shot 2017-03-24 at 9.04.14 AM




Traffic-Aware Cruise Control uses a camera

Mounted on the windshield behind the interior rear view mirror and a radar sensor in the
center of the front grill to detect whether there is a vehicle in front of you in the same
lane. If the area in front of car is clear, Traffic-Aware Cruise Control is designed to
drive consistently at a set speed.
When a vehicle is detected, Traffic-Aware Cruise Control is designed to slow down the car if
needed to maintain a selected time-based distance from the vehicle in front, up to the
set speed. Traffic-Aware Cruise Control does not eliminate the need to watch the road in
front of you and to apply the brakes if needed.
Traffic-Aware Cruise Control makes it easy to maintain a consistent time-based distance
from a vehicle travelling in front of you in the same lane. Traffic-Aware Cruise Control is
primarily intended for driving on dry, straight roads, such as highways and freeways. It
should not be used on city streets.

GOTURN tracking


Optical Flow experiments

Filtering noise with Kalman Filters

The workhorse of robotics is Kalman Filters


Training Neural Networks for classification using the Extended Kalman Filter: A comparative study (effen paywalled)

Speed control for safety

Note: bouncy image needs vectors stabilized by estimating image-based ego-motion estimation. Then you can measure the speed of objects in the driving window. Horizontal based ego-motions are probably lane dumpers. This estimation could predict lane dumpers and should affect speed and enhanced by other sensors.

  1. What is the distance to each of the objects
  2. What is the speed of the current trajectory?
  3. Optical flow for moving objects
  4. Remove bouncy image with prediction
  5. Filter stationary objects
  6. What objects are left to predict from?
  7. Calculate speed and directions of moving objects
  8. Which objects are travelling with me?
  9. What objects are travelling into my path?

FlowNet: Learning Optical Flow with Convolutional Networks

Optical Flow



A subject is to provide a pedestrian motion predicting device capable of accurately predicting a possibility of a rush out before a pedestrian actually begins to rush out. According to the embodiments, the pedestrian is detected from input image data, a portion in which the detected pedestrian is imaged is cut out from the image data, a shape of the pedestrian imaged in the cut-out partial image data is classified by collating the shape with a learning-finished identifier group or a pedestrian recognition template group, and the rush out of the pedestrian is predicted based on a result of the acquired classification.