Pedestrian Detection in Thermal Images by Enhancing the Brightness of Pedestrians

Thermal image, which is not affected by visibility, has attracted significant attention in pedestrian detection. How-ever, thermal images have poorer image quality and lack color and texture features compared with RGB images. Furthermore, pedestrian detectors based on deep learning often rely heavily on feature extraction networks. As a result, the performance of the detectors tends to decrease when directly applied to thermal images. To solve this problem, we design a pre-processing network to fully use the feature that pedestrians have higher brightness in thermal images. The pre-processing network can enhance the brightness of pedestrians in thermal images. Then we filter out the brightest area from the processed image by increasing the contrast, and input the filtered result into the detector together with the original image to help the detector find pedestrians. In addition, we found that an overly complex feature extraction network is redundant for thermal images and will have a negative impact. On this basis, we simplify the feature extraction network of YOLOv3. After simplification, the accuracy and running speed are improved, and the memory usage of the model is reduced. Through sufficient experiments on the KAIST dataset, it is proved that our method can significantly improve the performance of pedestrian detection.