Detecting and learning city intersection traffic contexts for autonomous vehicles

According to Business Wire, the global market for Autonomous Vehicles estimated at 6.1 thousand Units fpsin the year 2020, is projected to reach a revised size of 110.1 thousand Units by 2026, growing at a CAGR of 60.6% over the analysis period. This strong demand brings more research interests in autonomous vehicle systems. One of the hot topics is autonomous machine vision system and intelligent solutions. Most existing papers on autonomous vehicle machine vision systems apply machine learning models to address automatic object detection and classification issues to support automatic street traffic object detection and classification for vehicles, people/animal, traffic road signs, and signals. However, there is a lack of research results addressing automatic detection and classification of road contexts and transportation intersections under diverse weather conditions. In this work, we present an integrated machine learning model to address the issue and need in street intersection detection and classification based on road contexts and weather conditions. This paper reports our efforts in data collection, processing, and training based on existing data sets (such as BDD100k and COCO), and add a new training data set on street contexts and intersections (13 classes) and weather conditions (6 classes). This paper proposes a 2-stage integrated model to support the detection and classification of different types of traffic contexts and transportation intersection. In the first stage, two deep learning models (Yolo4 and Mask CNN) with transfer learning technique are used to detect the traffic signs, traffic lights, crosswalk and road direction targets on the road in addition to mobile objects (such as people and cars). Later, the generated results from the first stage are used as inputs to a decision tree model to detect and classify the different types of underlying transportation intersections. According to presented experimental results, the proposed 2-stage model receives a high accuracy, so it has strong potential application in computer vision technology of autonomous driving.