To compare video analytics solutions for a gas station, six criteria are sufficient to narrow down the options: compatibility with existing cameras, the actual use cases covered at this type of location, detection quality as measured on your own footage, the deployment and monitoring architecture across a network of locations, the compliance framework, and the total cost per location over three years.
These criteria carry more weight than the list of features in a sales brochure. A gas station typically includes a fueling lane, sometimes a car wash area, charging stations, a convenience store, and storage areas—all covered by a diverse array of cameras and operated by a very small team. A solution that requires replacing the entire camera system—or having an agent constantly monitoring a screen—will not scale effectively across a network of sites.
The needs fall into three categories, and not all solutions address them with the same level of sophistication.
Safety of people and property: detection of falls on the fuel island or in the store, intrusion after closing hours, prolonged presence near charging stations, parking in a restricted area, driving in the wrong direction on the fuel island.
Commercial metrics: counting visitors in the store, excluding staff from foot traffic counts, time spent in each zone, conversion rate between the fueling area and the store.
Operations: occupancy rates for charging lanes and stations, wait times, and the duration a charging spot remains idle after a charge is complete.
Ask each vendor to specify which of these use cases are available in production today, which are part of a roadmap, and which require a dedicated camera. The difference between the three answers often determines the final ranking.
This is the criterion that has the greatest impact on the budget. A software solution that integrates with existing video feeds—via a recorder or a standard network feed—avoids the need for a network-wide hardware replacement initiative.
Three questions to ask: Which protocols and stream formats are supported? What are the minimum resolution and frame rate requirements for each use case? And what processing is applied when a camera is misaligned or partially obstructed?
Also ask where the processing takes place. Analysis performed on equipment installed on-site limits the volume of images leaving the station. Remote analysis requires a network connection with sufficient bandwidth, which isn’t always available at remote locations.
The detection rates reported in sales materials are based on datasets you are not familiar with. They say nothing about how the algorithm will perform on your tracks, at night, in the rain, or with glare on wet concrete.
The only reliable method is to measure using your own images. Select two or three representative sites, including one challenging location. Establish a ground truth by annotating a real-world period—for example, two weeks—and then compare, for each candidate, the number of correct detections, the number of missed events, and the number of false alarms per camera and per day.
The number of false alerts per day deserves special attention. At a gas station, the on-duty staff is also serving customers. Beyond a certain volume, alerts are no longer addressed at all, regardless of how accurate the model is.
A successful pilot at one site does not guarantee a successful rollout across a hundred. The factors that make the difference are rarely algorithmic.
Check the procedure for adding a site: how long it takes, the skills required on-site, and whether an on-site visit is necessary. Check how remote updates are managed, how camera and processing status is monitored, and the ability to apply the same configuration to a group of sites without having to repeat the process site by site.
Finally, check the reporting. An operations network needs to compare its sites with one another and track trends, not just view a dashboard for each station.
A gas station is open to the public, and the service area is most often a space accessible to the public. The installation is therefore subject to the video surveillance regulations set forth in the Internal Security Code, requiring prefectural authorization and public notification via posted signs.
The addition of algorithmic processing to these images constitutes the processing of personal data within the meaning of the GDPR. It requires a documented legal basis, a defined retention period, and, in most cases, a data protection impact assessment, given the systematic surveillance of a publicly accessible area.
Since store employees are being filmed, they must be individually informed, and the Social and Economic Committee must be consulted before the system is put into operation. Systems based on facial recognition fall outside this framework and require their own legal basis, which explains why solutions intended for commercial use do not include them.
Ask each vendor for a document specifying what data is generated, which data is retained, for how long, and who has access to it. The absence of such a document is, in itself, a form of information.
Compare the cost per site per year over a three-year period, including the software license, any local processing hardware, installation, cameras to be added or repositioned, network connectivity, integration with your monitoring tools, and support.
Add one item that’s often overlooked: the time it takes for teams to process alerts. Two solutions with the same license price won’t have the same operating costs if one generates ten times as many notifications that need to be assessed.
XXII is a French computer vision software developer. Its CORE platform is designed to analyze video feeds from already-installed cameras—without facial recognition—and to address both security and operational measurement needs on a single platform, with centralized monitoring across multiple sites.
This framework remains the same regardless of which candidate is being evaluated. Its main advantage is that it shifts the discussion from a list of features to what the solution actually delivers for your images, your sites, and your teams.
First, choose the two or three use cases that justify the project, setting aside those that won’t have an operational beneficiary. Next, select the test sites, including at least one challenging site. Finally, establish quantifiable acceptance criteria before launching the tests—not after.
A comparison conducted in this order leads to a defensible decision. Conducted the other way around, it leads to a mere preference.