Dellaert’s article proposes a real-time model based vision approach for detecting and tracking vehicles from a moving platform. This approach tries to emulate situational awareness without making an assumption about the ground plane. This awareness is vital for any autonomous vehicle especially on the road with other cars. The image processing algorithm essentially takes the image of the road and treats it as a two dimensional grid. It scans for any edges of a car that stand out after a gradient is created from the image. These “edges” then form a bounding box around the car on the road. Bounding boxes are commonly used in two dimensional graphic displays so that there is a method for determining if another object is inside another object. This is a form of collision detection.
The primary issue that I see is that an image processing algorithm that is run in real time needs to be fast especially in the case of driving on the road where you have less than one second to respond to a sudden change, such as a car swerving into your lane or a deer in the middle of the road. I find it interesting that in a 3D environment, they can utilize a 2D solution where the 2D bounding box in the video stream, they can update the estimate of the 3D position of the tracked vehicle. The efficiency of this algorithm is the fact that it only has to process line segments which can be done extremely fast because the elements themselves are simple. Another issue with tracking this information in real time is if the algorithm can tie the subsequent frames of the video together to model the environment in real time. The tracking piece of this solution compares the bounding box in the previous frame with the bounding box in the current frame in order to see the change in the surrounding vehicles.
There are many variables to consider when dealing with real time road images, such as image noise in the video feed and weather conditions. As with any image processing, there needs to be an initial calibration in order to account for these variables. What I like about the solution to this approach is that they develop a probability model of the environment do determine the most likely candidates for the car outlines. The algorithm utilizes an extended Kalman-Bucy filter aids in the creation of a realistic model of how a tracked car behaves. This helps with giving a more accurate initial estimate to initialize the search for a bounding box, which will create better tracking performance.
Extending this approach, I would like to see them start to model how drivers actually drive on the road. Since probabilistic algorithms are already introduced to increase the performance of the overall solution, than I would recommend adding additional models to this space. I believe the lookup time would be minimal and this could significantly reduce the space in which the bounding box needs to be computed.
Reference:
- Dellaert, Frank. “Robust Car Tracking using Kalman filtering and Bayesian templates.” <http://www.cc.gatech.edu/~dellaert/pub/Dellaert97spie.pdf>
