Thermal sensors are underrepresented in the field of Advanced Driver Assistance Systems whereas their capabilities to acquire images independently of weather or daytime can be very helpful to achieve optimal pedestrian and vehicle detection. This underrepresentation is due to the small amount of available public datasets. This lack of training samples and the difficulties of building such datasets are a real hurdle to the development of an object detector dedicated to thermal images. Thanks to YOLOv4 and its detection performance, we show in this paper that finetuning this neural network requires few samples to achieve satisfying performance, outperforming the results of stateoftheart detectors.

Thermal images are widely used for various applications such as safety, surveillance, and Advanced Driver Assistance Systems (ADAS). However, these images typically have low contrast, blurred aspect, and low resolution, making it difficult to detect distant and small-sized objects. To address these issues, this paper explores various preprocessing algorithms to improve the performance of already trained object detection networks. Specifically, mathematical morphology is used to favor the detection of small bright objects, while deblurring and super-resolution techniques are employed to enhance the image quality. The Logarithmic Image Processing (LIP) framework is chosen to perform mathematical morphology, as it is consistent with the Human Visual System. The efficacy of the proposed algorithms is evaluated on the FLIR dataset, with a sub-base focused on images containing distant objects. The mean Average-Precision (mAP) score is computed to objectively evaluate the results, showing a significant improvement in the detection of small objects in thermal images using CNNs such as YOLOv4 and EfficientDet.