An Efficient Hybrid Residual Network for Pedestrian Detection and Segmentation

Authors

• Rahul Sawhney¹
• Aabha Malik¹
• Rishita Khurana¹

Abstract

Pedestrian detection is a vital and important issue which needs to be addressed as it has many applications in different fields such as advanced mechanics, automotive safety, and most importantly surveillance. A large part of the advancement of the previous few years has been driven by the accessibility of testing public datasets and proposing fruitful solutions to the problem. Object detection for the most part requires sliding-window classifiers in custom or anchor-based expectations in present day profound learning draws near. To proceed with the quick pace of advancement, another point of view where identifying objects is inspired as an undeniable level semantic object detection task is presented in this paper and further developed assessment measurements, showing that generally utilized per-window measures are less effective and can neglect to anticipate execution on full pictures are also presented. The proposed hybrid model has the baseline of MobileNet, with the integration of skip connection of ResNet50 with Faster RCNN and focuses all over the picture and extract all the important features that are required for detection. Nonetheless, in contrast to these customary low-level provisions, the proposed model goes for a more elevated level detection. Subsequently, in this paper, pedestrian detection is streamlined as a straight-forward focus and scale expectation task through convolutions. Thusly, the proposed technique is basically basic, it presents competitive precision and great speed and prompts a new attractive pedestrian detector.

Dataset

https://www.cis.upenn.edu/~jshi/ped_html/