A novel pedestrian detection and tracking with boosted HOG classifiers and Kalman filter

This paper focuses on developing a stable pedestrian detection and tracking algorithm. Although Histogram of Oriented Gradients (HOG) features are the best representation for human shapes, computing these feature vectors are computationally expensive as it slows down the overall detection process. Hence with the use of cascade of boosted classifiers, the overall process was shortened significantly even in the absence of graphics processing unit (GPU). Along with Kalman filter approach, the algorithm achieved desirable results in tracking pedestrians coming from various directions. The Kalman filter model with its self-correcting mechanism, guarantees that the tracking improves overtime as more raw detections are supplied. As long as consistent detections were supplied to the filter in the early stages, the tracking continues even when the detector becomes faulty.