Contact person: AI Tool analyzing satellite imagery to detect a missing vehicle

Phone:Show

Email:Show

Location: Toronto, Canada

Budget: Recommended by industry experts

Time to start: As soon as possible

Project description:
"I want a straightforward AI solution that scans publicly-available high resolution satellite imagery (think Google Earth tiles, Sentinel, Landsat, etc.) and flags anything that looks like the missing vehicle I am searching for. The vehicle is a black SUV approximately 178" long, 72" wide, 66" high. The car will have been in an accident, and will be off a road or in water. Trees, as well as buildings with larger dimensions than the vehicle, and roads and other infrastructure, can be excepted from the search

I am looking for object detection. The vehicle will have been in the same place for almost three years. Change-over-time analysis can be used to eliminate objects which have moved over time. If the car has moved during the last two years, it is not the car I am looking for.

Cars that are on roads or in driveways can be eliminated from the search. They will also have moved over time and can be eliminated.

This is a confidential, work for hire engagement. I require a transfer of all IP rights in the solution and results to me.

Because the satellite images are coarse and often outdated, I am realistic about accuracy—anything under 75 % confidence is acceptable as long as I receive a ranked list of potential hits I can manually review. Speed of processing and an easy way to retrain or tweak the model matter more to me than perfect precision.

The search area is a known area - I know where the vehicle left from, and where it was going, less than 10 kilometres away. The search radius can begin within the first 50km from the site of disappearance, and move outward to no more than 150km.

Here is what I need delivered:
• A Python-based script or small app (TensorFlow, PyTorch, YOLO, or a comparable open-source framework is fine) that ingests raw satellite tiles, runs object detection, and saves bounding boxes plus confidence scores to a simple CSV/GeoJSON.
• Clear setup instructions and a brief read-me so I can point the tool at new imagery without touching the code.
• One run on sample public data (preferably Google Earth) demonstrating the detections within a 50km radius, and one run on a 150km radius.

Feel free to leverage pre-trained weights or lightweight transfer learning; I just need a working prototype that helps narrow my manual search grid quickly." (client-provided description)