Night Pedestrian Detection Using Thermal Image Feature Extraction Enhanced YOLOv3 (TIFEEY)

False identification of pedestrian detection in autonomous driving can potentially lead to traffic accidents. The traditional visible light detection method is difficult to detect accurately under low illumination and fog conditions. To address these issues, a nighttime pedestrian detection model with infrared thermal imaging technology combined with deep learning algorithms is proposed, which is called Thermal Image Feature Extraction Enhanced YOLOv3(TIFEEY). Specifically, for nighttime infrared pedestrian images, the pedestrian candidate regions are divided, then the images are enhanced by contrast stretching. In addition, a novel downsampling structure containing convolution and maximum pooling is used to obtain feature layers, then a convolutional layer containing multiple asymmetric convolutions is used to extract pedestrian features, and finally YOLOv3 is used to regress and classify the features. However, the network complexity limits the detection speed, so the number of feature layers is pruned by channel pruning. The FLIR dataset is constructed and divided into a training set and a test set, and the model achieved a detection accuracy of 90.94% on the test set, with an inference time of only 22ms per image. In summary, it demonstrates that the proposed TIFEEY model can quickly and accurately detect pedestrians at night. Therefore, it can reduce the risk of using solely visible light detection at nighttime.