Frequently Asked Questions
Troubleshooting tips and tricks
What about privacy?
The Traffic Monitor is built with a privacy-by-design approach because we believe in:
Ethical and Transparent Monitoring: Our mission is to democratize smart city monitoring by making it ethical and transparent. We value transparency from open source, glass-box code, architecture, and models.
Community Safety and Trust: The architecture is designed to maintain privacy and keep the population safe while providing data in your control. We understand it is ultimately in the hands of the person implementing a solution. A solution that is open can be better understood by anyone, and audit controls may be demanded from operators, especially government departments.
Unbiased Data Collection: Object detection and counting is performed without human bias, allowing for a whole new avenue to collect data on holistic road usage. We understand and acknowledge that algorithmic bias may be built into the machine learning models intentionally or unintentionally. To address this we use public domain and open source models whenever possible and may specify models for full transparency.
How is privacy implemented in the Traffic Monitor?
Privacy is implemented through a decentralized, local processing architecture:
Local Inferencing (Edge ML): All image processing, model inferences, and initial data storage occur locally on the device. This means no cloud data processors or third parties are involved in the core data processing.
Native Data De-identification: The device only records non-uniquely identifiable metadata (such as object label, speed, and timestamps) to an internal database, unless additional models are enabled such as License Plate Recognition (LPR).
No Images or Videos Stored by Default: The native data de-identification process ensures no images or videos are stored unless the monitor is specifically configured for surveillance purposes (e.g., for License Plate Recognition or to capture a speeding event).
User Control Over Sharing: Users control what data is shared from the device through open communication standards like MQTT and HTTP APIs.
Open Source for Transparency: The system's software is open source (OSS) for transparency and auditability.
Privacy Regions: The device may be configured with privacy regions to restrict what data is collected in certain areas, see Defining Masks.
How many simultaneous object detections can the traffic monitor handle?
I am interested in using this in a location that is very congested. There could easily be 10 walkers and a few bikes in a single frame. Can it handle that many simultaneous detections? I am primarily looking for counts and direction of travel.
Short answer
Yes, 👍 it can handle all that and more than you will throw at it. I think the biggest practical limitation for the Traffic Monitor is likely going to be when you have so many objects they "overlap" each other, so you can't tell if there is a person/bike behind another (think at the beginning of a marathon).
Longer answer
We use the Frigate NVR Project to do the heavy lifting on object detection, decoding video, and motion detection. So, I am going to reference their documentation and discussions.
A couple of relevant discussions:
if you have an inference speed of 10 milliseconds then that means you can run (1000 ms in a second / 10 milliseconds per inference) 100 object detection inferences per second. Frigate often runs multiple inferences on a single camera frame, for instance when motion happens in two places at the same time, when an object was partially detected at the edge of a region and a larger region should be used, etc.
from https://github.com/blakeblackshear/frigate/discussions/7491 (emphasis mine)
The maximum number of detections per frame is determined by the largest number of disjoint movement zones, for which the upper bound then is equal to how many tiles of WxH, where WxH is model input dimensions, are needed to tile the full frame (not counting the region-connecting logic).
Some illustration of this. I have a quiet road today, but when the sun is casting shadows through the trees, Frigate is doing a lot of work to send various regions (green boxes) to object detection to "see" if any of that shadow motion is an object. The following image has more than 18 regions (that I can count) it is sending to object detection PLUS the 9(??) labeled objects it is tracking and sending. My inference speed was still sitting around 9ms, so the object detector could handle many more. The CPU spikes because of the video decoding and motion tracking, but with the RPi5 we still had quite a bit of overhead.
Last updated
Was this helpful?