Improving Street Object Detection Using Transfer Learning: From Generic Model to Specific Model

抄録

A high accuracy rate of street objects detection is significant in realizing intelligent vehicles. Algorithms based on convolution neural network (CNN) currently exhibit reasonable performance in general object detection. For example SSD and YOLO can detect a wide variety of objects in 2D images in real time; however the performance is not sufficient for street objects detection, especially in complex urban street environments. In this study, instead of proposing and training a new CNN model, we use transfer learning methods to enable our specific model to learn from a generic CNN model to achieve good performance. The transfer learning methods include fine-tuning the pretrained CNN model with a self-made dataset, and adjusting the CNN model structure. We analyze the transfer learning results based on fine-tuning SSD with self-made datasets. The experimental results based on the transfer learning method show that the proposed method is effective.