New Features and Insights for Pedestrian Detection Stefan Walk1 Nikodem Majer1 Konrad Schindler1 Bernt Schiele1,2 1 Computer Science Department, TU Darmstadt 2 MPI Informatics, Saarbrücken Abstract Despite impressive progress in people detection the per- formance on challenging datasets like Caltech Pedestrians or TUD-Brussels is still unsatisfactory. In this work we show that motion features derived from optic flow yield sub- stantial improvements on image sequences, if implemented correctly—even in the case of low-quality video and conse- quently degraded flow fields. Furthermore, we introduce a new feature, self-similarity on color channels, which con- sistently improves detection performance both for static im- ages and for video sequences, across different datasets. In combination with HOG, these two features outperform the state-of-the-art by up to 20%. Finally, we report two insights concerning detector evaluations, which apply to classifier-based object detect
