Improving pedestrian segmentation using region proposal-based CNN semantic segmentation

Abstract

Pedestrian segmentation is a critical task in computer vision, but it can be challenging for segmentation models to accurately classify pedestrians in images with challenging backgrounds and luminosity changes, as well as occlusions. This challenge is further compounded for compressed models that were designed to deal with the high computational demands of deep neural networks. To address these challenges, we propose a novel approach that integrates a region proposal-based framework into the segmentation process. To evaluate the performance of the proposed framework, we conduct experiments on the PASCAL VOC dataset, which presents challenging backgrounds. We use two different segmentation models, UNet and SqueezeUNet, to evaluate the impact of region proposals on segmentation performance. Our experiments show that the incorporation of region proposals significantly improves segmentation accuracy and reduces false positive pixels in the background, leading to better overall performance. Specifically, the SqueezeUNet model achieves a mean Intersection over Union (mIoU) of 0.682, which is a 12% improvement over the baseline SqueezeUNet model without region proposals. Similarly, the UNet model achieves a mIoU of 0.678, which is a 13% improvement over the baseline UNet model without region proposals.