Robust pedestrian detection using scale and illumination invariant Mask R-CNN

In this paper, we address the challenging difficulty of detecting pedestrians with variation in scale and the illumination of the images. Occurrences of pedestrians with such variations exhibit diverse features. Therefore, it intensely affects the performance of recent pedestrian detection methods. We propose a new robust approach for overcoming the antecedent challenges. We proposed a scale and illumination invariant Mask R-CNN (SII Mask-RCNN) framework. The first phase of the proposed framework corrects illumination variations by performing a logarithmic transformation and adaptive illumination enhancement. In addition, the non-subsampled contourlet transform is used to decompose the image into multi-resolution components. Finally, we convolved the image with the multi-scale masks to find corresponding points that are illumination and scale-invariant. Extensive evaluations on pedestrian benchmark databases illustrate the effectiveness and robustness of the proposed framework. The experimental results contribute the notable performance improvements in pedestrian detection as compared to the state-of-the-art approaches.