Abstract
Fast and robust video based pedestrian detection has been an active research area in computer vision for the past many years and is still considered a challenging task. Despite appearance of several sophisticated algorithms performing conspicuously on standard datasets, the goal of achieving an acceptable detection performance in real-time scenario remains elusive. Earlier works in this regard have demonstrated that kernel support vector machine based classifiers can be integrated with low complexity integer-only features to yield powerful detectors achieving accuracy at par with the contemporary boosting cascades based detectors which employ complex features requiring floating point operations. The current work describes a new technique to implement a soft cascade to speed up evaluation of the kernel classifier. The proposed approach achieves rapid early rejection of the true negatives through identification of the most relevant feature components sorted by the estimated energies of their corresponding kernel functions. The proposed cascading scheme ensures that evaluation of only 4% of the features in the detection window leads to rejection of up to 62.8% and 61.8% true negatives on INRIA and ETH datasets respectively without noticeable effect on detection accuracy. Overall, a reduction of 90% and 89% in the computational load of kernel evaluation is observed on the mentioned datasets respectively while showing excellent generalization property. This technique combined with inclusion of multiple detectors for different scales and hardware support for parallelization and vector computation results in up to three times the processing speed on desktop as well as embedded processors compared to the contemporary detectors while achieving better or comparative detection accuracy.