Pedestrian Detection Algorithm in SSD Infrared Image Based on Transfer Learning

Because of difficulty in feature extraction of infrared pedestrian images, the traditional methods of object detection usually make use of the labor to obtain pedestrian features, which suffer from the low-accuracy problem. With the rapid development of machine vision, deep learning has gradually become a research hotspot and a mainstream method for many pattern recognition and object detection problems. In this paper, aiming at the defects of deep convolutional neural network, such as the high cost on training time and slow convergence, a new algorithm of SSD infrared image pedestrian detection based on transfer learning is proposed, which adopts a transfer learning method and the Adam optimization algorithm to accelerate network convergence. For the experiments, we augmented the OUS thermal infrared pedestrian dataset and our solution enjoys a higher mAP of 94.8% on the test dataset. After experimental demonstration, our proposed method has the characteristics of fast convergence, high detection accuracy and short detection time.