基於深度學習之車輛前方障礙物距離估測

Distance Estimation of Obstacles in Front of Vehicles Based on Deep Learning

由於科技的進步，自主駕駛系統在可知的未來必形成一股風潮，而車輛與障礙物之間的距離估測是自主駕駛系統中一個非常重要的技術。為了達到距離估測之目的，目前發展的自主駕駛系統大都需要仰賴各式各樣的距離感測器，例如光達、雷達及超音波等，這些感測器在距離量測上通常具有高精度，但同時也伴隨著高昂價格，這將使得自主駕駛系統的推廣及普及變得愈發困難。本文提出了一個結合語義分割與深度估測之深度神經網路模型，其包含有相同卷積層數的Encoder與Decoder網路。本文所提之網路架構在KITTI及Cityscapes資料集上進行訓練，並在最後結合語義分割與深度估測等方法進行距離估測，實驗結果證實本文所提方法的可行性。

Autonomous driving systems are the wave of the future; for such systems, the estimation of the distance between the vehicle and surrounding obstacles is key. Most current distance estimation methods rely on a variety of distance sensors, such as LiDAR, radar, or ultrasonic sensors. Although these sensors measure distance accurately, their high cost hinders the popularization of autonomous driving systems. To remedy this problem, this paper proposes a deep neural network (DNN) that combines semantic segmentation and depth estimation. The DNN includes an encoder and a decoder, both of which have the same number of convolutional layers. The proposed network architecture was trained on both the KITTI and Cityscapes datasets. The proposed method provided accurate distance estimation in evaluation tests, demonstrating its feasibility.