Pedestrian Localization
Markus Gressmann, Otto Löhlein, Günther Palm · 2011
This work investigates the problem of precise localization of pedestrians in images. The difference between pedestrian detection and localization is illustrated and it is shown how the localization task can be cast as a ranking problem. A novel Ranking Neural Net using Local Receptive Fields that operate directly on pixel values is proposed. The feature extraction layer of the network can be trained to fit the underlying pixel data, which makes it responsive to even small shifts in position or scale. To find the most likely position of a pedestrian, the network can be applied to the image in either an exhaustive fashion, or very efficiently using gradient descent. The performance of the proposed network architecture is evaluated on images of the publicly available Daimler Pedestrian Detection Benchmark and compared to a standard detection approach.