Brigit A. Schroeder
December 12, 2011

The goal of the this project was to use computer vision-based object tracking and probabilistic recognition methods to automatically identify pedestrians in fixed surveillance video.

Concepts Demonstrated

  • Several computer vision concepts are demonstrated in the moving object tracker, such as frame differencing, image histograms, and connected components analysis, which is used to segment out targets for classification.
  • Histograms of Oriented Gradients are used as the primary input features for pedestrian and background target classification. The classifiers have been trained and tested using the Naive Bayes classifier with log-likelihoods.


Tracking and classification are two fundamental topics in computer vision and AI. Putting both together in a functioning pipeline without crushing computational burden and with accurate classification are some of the greatest challenges. In this project I introduce a summarized form of Dalaal and Triggs (CVPR, 2005) histogram of oriented gradients to reduce the the classification feature vector size and thus the computational burden of processing a normally large feature vector. I also modified the original proposed tracker to incorporate a simple background model for frame differencing to detect motion in an image.

Evaluation of Results

The tracker works very well for fixed surveillance video. The summarized histogram of oriented gradients method I introduced in conduction with Naive Bayes has a > 90% detection rate on pedestrians in the training sets I used (INRIA, MIT Pedestrian Set) and a < 5% false alarm rate. However, the classification of targets tracked and segmented from surveillance video proved to be more due issue of target scale, differing quality/sensitivity of surveillance cameras and possibly the need for a more robust/diverse pedestrian training set.

Tracker Target Classification Example

Additional Remarks

Future work (e.g. more time) will be needed to examine the multi-scale classification problem.