Edge Enhanced Traffic Scene Segmentation Algorithm with Deep Neural Network

摘要

Image segmentation is critical in autonomous driving field. It can reveal essential clues such as objects’ shape or boundary information. The information, moreover, can be leveraged as input information of other tasks: vehicle detection, for example, or vehicle trajectory prediction. SegNet, one deep learning based segmentation model proposed by Cambridge, has been a public baseline for scene perception tasks. It, however, suffers an accuracy deficiency in objects marginal area. Segmentation of this area is very challenging with current models. To alleviate the problem, in this paper, we propose one edge enhanced deep learning based model. Specifically, we first introduced one simple, yet effective Artificial Interfering Mechanism (AIM) which feeds segmentation model manual extracted key features. We argue this mechanism possesses the ability to enhance essential features extraction and hence, ameliorate the model performance. Other modifications of model structures were also designed for further improving model’s feature extraction ability. Besides, one Pixel Alignment Unit (PAU) was presented for pixel level alignment. The unit is designed based on Bidirectional Long Short Term Memory (Bi-LSTM) unit and, according on our design, is able to reconstruct and extract pixel spatial features which is a key clue for the segmentation. Combined with mentioned methods, in the end, an integrated model was proposed. To evaluate our model, CamVid dataset were adopted in experiments. The experiment result showed that our model has the ability to refine objects margin area segmentation results. Our contribution lies in that we attempt to boost model performance through artificially interfered model feature extraction phases and attempt to adopt the Bi-LSTM structure to reconstruct and extract pixels’ spatial features.