Wide Area Car Detection

Model by Orbital Insight

Orbital Insight’s car detection algorithm produces point detections corresponding to the location of cars within a specified input image. Subscribers can be alerted of when there is a significant increase or decrease in car traffic, and view a historical dataset to create a baseline from. This saves analysts significant amounts of time in conducting pattern of life analyses and activity based intelligence. It can be used for deriving economic indicators, observing deployment of public services, infrastructure, retail and manufacturing sites and monitoring conflict zones.

  • Description

    Product Description

    PERFORMANCE METRICS

    The algorithm has a precision of 0.91 and a recall of 0.89. As this is an overall average representing the algorithm’s performance, its performance will vary according to geography and imagery features. For example, the algorithm is more reliable in urban areas. The algorithm generally performs better in North American and European areas of interest. The algorithm has known limitations when dealing with highly shadowed imagery, those containing closely packed vehicles (0 pixels in between), and desert imagery.

    OVERVIEW:

    Orbital Insight’s car algorithm automatically detects and quantifies cars from daily satellite imagery using multiple commercial providers. The algorithm’s computer vision is able to accurately, economically, and efficiently count objects at scale yielding economic and geopolitical insights. Here are just a few of our supported use cases:

    Patterns of Life / Impact Assessment Commercial and passenger traffic through Marawi were severely disrupted due to a battle between government forces and insurgents in May 2017. Citywide car counts immediately declined once the battle started, and those in the most heavily affected regions have not completely recovered to their pre-battle numbers.

    Economic Development Addis Ababa is one of the fastest growing cities population-wise in the world. As the Chinese have been investing heavily in the area, particularly some of those smaller areas on the periphery of the city, we can demonstrate the upward trend in traffic patterns compared to prior years.

    Places of Interest Analysis Car counts serve as a human-activity metric within and across well-defined places of interest, providing insight into activities across multiple domains. Interrelated locations can be monitored in large numbers, and across long periods of time, to automatically detect trends and anomalies. In this example, we see disruption to commercial air traffic in Damascus, Syria, in both a pre- vs. during conflict comparison as well as a long term trendline; these findings reflect the impact of that country’s civil conflict on society and local economies. This methodology is transferable to any well-bounded set of facilities, and the impact of findings is reflective of the function of said locations.

    TRAINING:

    The training set contains 180,000 images and spans many variable conditions including time of day, time of year, terrain, configuration of vehicles, etc.

    VALIDATION:

    This algorithm was tested by a 3rd party using a validation set, consisting of approximately 100,000 marked cars in 50 countries spanning 20,000 images of deserts, ports, and parking lots across 6 continents.