Effects of Pre-processing on the Performance of Transfer Learning Based Person Detection in Thermal Images

Thermal images have the property of identifying objects even in low light conditions. However, person detection in thermal is tricky, due to varying person representations depending upon the surrounding temperature. Three major polarities are commonly observed in these representations i.e., 1. person warmer than the background, 2. person colder than the background and 3. person’s body temperature is similar to background. In this work, we have studied and analyzed the performance of the detection network by using the data in its original form and by harmonizing the person representation in two ways i.e., dark persons in the light background and light persons in a darker background. The data passed to each testing scenario was first pre-processed using histogram stretching to enhance the contrast. The work also presents the method to separate the three kinds of images from thermal data. The analysis is performed on publicly available outdoor AAUPD-T and OSU-T datasets. Precision, recall, and F1 score is used to evaluate network performance. The results have shown that network performance is not enhanced by performing the mentioned pre-processing. Best results are obtained by using the data in its original form.