Pedestrian Detection: Fid-Bau Portal

Pedestrian Detection

Deshpande, Shashank / Cai, Yang

Pedestrian detection is critical to intelligent transportation systems, ranging from autonomous driving to infrastructure surveillance, traffic management, transit safety and efficiency, and law enforcement. Pedestrian detection involves many types of sensors, such as closed‐circuit television cameras (CCTV), thermal imaging devices, near‐infrared imaging devices, and on‐board RGB cameras. There is also a broad spectrum of pedestrian detection algorithms based on infrared signatures, shape features, gradient features, machine learning, or motion features. In this chapter, we focus on exploring the methods for pedestrian detection and recent developments using Thermal imaging, near‐infrared light imaging, background modeling, shape analysis, and multi‐sensor fusion. We also address the relationship between these algorithms.

For CCTV camera applications, we find that Approximated Median is more accurate than Frame Subtraction and faster than the Gaussian Mixture Model. Our empirical results show that Approximated Median performs well in detecting pedestrians in outdoor and indoor environments. The Polar Coordinate Profile can map the 2D shape description to one‐dimensional models, which simplifies pattern‐matching algorithms. HOG enables individual image‐based detection, based on the “cell‐like” histogram of oriented gradients. However, it has several limitations, including issues with scalability, occlusion, deformation, unconstrained backgrounds, complex illumination patterns, and a wide variety of articulated poses and clothing. The deformation problem can be solved by applying a Deformation Parts Model. However, occlusion handling is a serious problem for pedestrian detection algorithms. LiDAR‐based method can reduce the false positives and computational time for pedestrian detection. By fusing LiDAR and visual camera data, we can improve the reliability and accuracy of the overall detection rate.