Pedestrian Detection Based on Center, Temperature, Scale and Ratio Prediction in Thermal Imagery

Abstract

Pedestrian detection has become an important branch in the field of object detection. However, pedestrian detection algorithms based on visible imagery meet many challenges under low lighting or no lighting conditions. Thermal imagery can effectively compensate the information loss of visible imagery in the night environment. In order to solve the problem of pedestrian detection at night or in other bad conditions with limited vision, a pedestrian detection method based on center, temperature, scale and ratio prediction in thermal imagery is proposed. The proposed method contains two parts: (1) feature extraction part; and (2) center, temperature, scale and ratio prediction part. The feature extraction part uses ResNet-101 network to extract high-level semantic features from input thermal imagery. In the center, temperature, scale and ratio prediction part, the center branch is to determine whether there is a target center point in each position of feature map, which is a binary classification problem. The temperature branch is to determine whether the center point is a pedestrian target radiating heat or the background. Scale prediction branch and aspect ratio prediction branch are to determine the size of the target, which is a regression problem. We test the proposed detection method in FLIR advanced driver assistance system (ADAS) dataset. Experimental results show that the proposed method can effectively deal with pedestrian detection in low light and dark environment, and its detection performance is better than that of the baseline of nighttime pedestrian detection methods.