March 10, Greenbelt Road, Suite 400, Greenbelt, MD Tel: (301) Fax: (301)

Transcription

1 Detection of High Risk Intersections Using Synthetic Machine Vision John Alesse, Brian O Donnell, brian.odonnell.ctr@dot.gov Stinger Ghaffarian Technologies, Inc. Cambridge, Massachusetts March 10, 2017 Abstract SGT has developed software and data analytics to analyze thousands of police accident reports. During a classification operation, contributing factors to an accident were subdivided into two major categories; human changeable such as roadway, obstructions, lighting and signage, and acts of nature such as weather, time of day and season. One particular area under human changeable factors involves visible obstructions. Accidents involving a stop sign at an intersection of a road where visibility was obstructed by a building were investigated (see Figure 1). In situations where the traffic on the cross road has unrestricted movement (no stop sign, yield sign, or traffic light) we can see significant increase in risk. The motorist at the stop sign would always be in the high risk position, when traffic was present, to enter the cross road as his view of the road would be occluded by the building. Figure 1: Satellite image from Google Maps of an intersection that has obstructed views for motorists headed south on South Congress Street and entering 6 th Avenue SE (Aberdeen, SD) 1

2 Automated Machine Vision Robust computer vision solutions for real-world scenery are abundant today as a result of millions of dollars spent on research on autonomous vehicles. These solutions typically require massive amounts of training data (i.e. big data ), advanced machine learning methods including deep learning which require clusters of high performance computers. By reducing the data representation from complex full-color world scenery, such as aerial or satellite photographs, to a finite colorized block diagram, the problem becomes quite simple. The simplified feature representation lends itself well to conventional machine vision. We refer to this as synthetic machine vision. What is Machine Vision? Machine vision is the use computer vision for industrial applications typically for the automated inspection of products. Usually the purpose is for assisting machines in the assembly of parts or for quality assurance. These applications, unlike problems with driver-less vehicles, are typically very constrained, well controlled, and require robust and high reliable results. Although one would expect a driver-less vehicle to require robust results there are many other sensors such as ultrasound, radar, and LiDAR to help form intelligent decisions. In machine vision the output is one of the following: geometric measurements, presence or absence of a part or feature, or a quality metric. Two examples of machine vision applications (see Figure 2) include inspecting the print on pharmaceutical products and providing vision for assembly robots. Figure 2: Machine Vision applications. Inspection of pharmaceutical products for quality of print (Left), assisting automotive manufacturing robots. 2

3 Google Maps Google Maps offers two standard map views: a full color, good resolution, satellite image, and a road map version. The road map version is provided with labeled roads of various types, labels and icons for places of interest (e.g. for businesses, government buildings, public institutions, etc.), and outlines of houses, buildings and other structures. In Figure 3 and Figure 4 the Google Maps views of buildings and streets are very simple in color palettes and structure. This lends itself well to our synthetic machine vision approach. Figure 3: Examples of building: residential (Left), commercial (Middle), and government (Right) Figure 4: Examples of streets. Regular roads (Left) are white and high traffic capacity/major roads such as highways are yellow and freeways orange (Right). Figure 5 shows the Google map color palette which can be used to create a look-up-table or dictionary for coding a machine vision solution. 3

4 Figure 5 Google Map s color scheme [1] Lack of Metadata One of the biggest challenges to the project was finding a source for roadway, signage and building geolocation metadata. Google has logged millions of miles of street view data, however the geolocation of roadway markings signs and surroundings was not captured. At the time of this writing, there is no known single source for geolocations of stop signs and buildings throughout the U.S. For the proof of concept, we were able to obtain a comprehensive list of stop sign locations from the Tennessee Roadway Information System (etrims) [2] (see Figure 6). Figure 6: Geolocations of stop signs from etrims [2] 4

5 Solution Using stop signs as known reference points along with the footprint building images from a basic Google mapped area, a solution was developed that can determine where obstructions occur thereby determining if the intersection was at high risk due to an obstruction. Python toolkits and the open source computer vision and machine learning software library Open Source Computer Vision Library (OpenCV) was used for importing images, image processing, and the analysis of edges/transitions to determine the boundaries of streets and buildings. OpenCV is a powerful library written in C/C++ and designed for fast processing. It provides support for CUDA and OpenCL interfaces and has interfaces to C++, C, Python and Java. It supports Windows, Linus, Max OS, ios, and Android. The NumPy library, a Python library for scientific computing, was used for fast mathematical operations. Image processing algorithms can determine the boundaries and direction of the roads. The Google Maps API also provides the direction of traffic (for example, N, S, E, W). Figure 7: Scanning the field of view for boundaries and occlusions. 5

6 Police Accident Report Accident Lat. & Long. US DOT Database Stop Line Lat. & Long. Synthetic Machine Vision Algorithm Occlusion Report Google Maps Google Map with Scale Factor Figure 8: Occlusion risk analysis algorithm. Figure 9: Synthetic Machine Vision Algorithm References [1] Mark Li, Zhou Bailiang, Discover the action around you with the updated Google Maps, Google, July 25, 2016, blog.google/products/maps/discover-action-around-you-with-updated/, Accessed March 5, [2] Tennessee Roadway Information System etrims, 6