Click on any picture to play video. If you have trouble playing any video, it is probably because your system does not already have the proper codec installed. To solve, you should either use the VLC player or install the full K-Lite codec pack, both of which are free.
Color code: cars (green), trucks (red), motorcycles (blue), motorcycle pairs (yellow). Note that the results are shown in real time, without any postprocessing cleanup. As a result, many of the brief rectangle flashes do not translate into actual counting errors; rather they display the "best guess so far" but are removed internally after subsequent image frames are processed.
|Middle of the road|
|More at night|
|More two-lane highway|
|Incident detection: stopped vehicle|
|Incident detection: wrong-way vehicle|
|Incident detection: pedestrian in the road|
|Automatic recalibration after pan/tilt|
As of May 2008, we have developed another algorithm with high-accuracy motorcycle, car, and truck capabilities:
Results of the boosted cascade pattern detector algorithm:
Note that the detector was not trained on the sequence presented but rather on different sequences with different conditions, angle, etc. The detector appears to work well even when there is non-zero roll angle. The results show simply the instantaneous detection results; temporal filtering will improve the results even further.
Results of the feature tracking algorithm on several sequences:
Vehicles are detected and tracked in the zone outlined by the red box. The inset shows the vehicle ID, classification as truck or car (T means truck), lane number, and speed (-1 means not enough data available). Note that lane numbers are reversed, i.e., Lane 1 is the slowest lane. Algorithm is based on grouping features and is described in IEEE Transactions on Intelligent Transportation Systems, 2008.