Pedestrian Detection for Autonomous Cars: Occlusion Handling by Classifying Body Parts
Muhammad Mobaidul Islam, Abdullah Al Redwan Newaz, Balakrishna Gokaraju, Ali Karimoddini
Abstract
In this work, we address the problem of detecting body parts of pedestrians using deep neural networks. In particular, we consider the occluded pedestrian detection problem in autonomous driving settings. While state-of-the-art deep neural models perform reasonably well for detecting full-body pedestrians, their performances are not satisfactory for occluded pedestrians. Introducing a new training strategy along with a fusion mechanism, we enhance the performance of the SSD-Mobilenet and the Faster R-CNN by utilizing body parts information to handle occluded pedestrians. We evaluate our method by training these two deep neural networks using a public dataset as well as our dataset. The performance of the two developed models is compared both in terms of detection accuracy and runtime efficiency.