With the rapid development of autonomous driving technology, obstacle recognition and range measurement for intelligent vehicles have become critical research areas. This study aims to delve into the data collection, preprocessing, and annotation processes of the obstacle recognition model for intelligent vehicles. Additionally, it explores the utilization of various deep learning models, such as the FCN and DeepLabV2networks, to achieve accurate obstacle identification. Furthermore, we investigated distortion coefficient solving and calibration methods for binocular cameras to achieve accurate measurement of obstacle distances. This paper conducted experiments using the ACDC dataset. Through extensive experimentation and analysis, we found that DeepLabV2achieved a recognition accuracy as high as 0.839, with a detection speed of 0.539 images per second. Additionally, we provided a comparative table of recognition accuracy for each model to assist researchers and engineers in better understanding the performance of these models. The calibration method proposed in this paper for binocular cameras has an error proportion lower than 2%. The achievements of this study are expected to propel further advancements in intelligent vehicle technology, enhancing road safety and the feasibility of autonomous driving.