We select the camera and lighting, prepare the image data, train the model and run it on edge hardware. You receive a working bench prototype, test results and specifications for your machine builder or integrator.
01Camera + lighting tested
02Model errors documented
03Runtime benchmarked
01 / Use cases
We test each task with representative images or video and define how the result will be checked.
We test whether the system can find missing parts, wrong orientation, print differences and visible surface defects under a defined camera and lighting setup.
We follow selected objects through video and return positions, trajectories, counts or timestamped events. This can cover material flow, equipment movement or sports footage.
We convert images or video into timestamped positions, state changes or calibrated dimension estimates. The measurement conditions and expected uncertainty are documented.
We build engineering prototypes for industrial, medical R&D and sports applications. We do not provide certified medical devices or safety-rated machine controls.
02 / Completed project
This bachelor’s project covers training data, model development, TensorRT acceleration and a browser interface for checking every saved ball position against the source video.
Bachelor’s thesis · working system
The system processes single-camera tennis footage and stores every detected position with its original video-frame number. This makes missed and incorrect detections easy to review.
Wide search · MobileNetV4 vs TrackNetV2
4.45×higher raw throughput · F1 0.895 vs 0.893Local tracking · MobileNetV4 vs TrackNetV2
3.45×higher raw throughput · F1 0.947 vs 0.950Accuracy uses 3,090 heatmaps from games 30 and 31, which were not used for training. Throughput is batch-one, GPU-resident TensorRT FP16 on an RTX 4060—not complete video-processing speed.
A player detector defines a wider search area when the ball is lost. After detection, a smaller crop follows the ball frame by frame. A lightweight TrackNet variant runs with TensorRT, and each saved position includes its original video-frame number.
03 / What we build
We develop these parts as one prototype, because image quality affects model accuracy and processing speed.
Which setup shows the required feature clearly?
We test camera type, lens, distance, angle, exposure and lighting on representative good and bad samples.
Does the model handle the agreed cases?
We prepare and label the image data, train or adapt the model, then report missed cases, false alarms and difficult examples.
Does it run fast enough on the target hardware?
We run the model on NVIDIA Jetson, an industrial PC or your server, measure processing time and connect the output to a prototype API or I/O interface.
Local processing · NVIDIA Jetson
We deploy the prototype model on a Jetson development kit and measure processing time using the intended camera resolution and workload. You receive the benchmark and a compute-hardware recommendation.
04 / How we work
We agree on the test cases, build the camera, model and hardware setup, then evaluate it on images or video that were not used to train the model.
We list the good parts, defects, objects or events to test; the required output; cycle time; and the errors the process can tolerate.
We test camera, lens and lighting, prepare the image data, train or adapt the model and run it on the selected hardware.
We run the agreed tests and deliver the prototype, test report, hardware specifications and interface documentation.
Built by Buchendorfer Vision
We deliver the capture setup, camera and lighting recommendation, trained model, edge benchmark, prototype software interface and the code and documentation agreed in the scope.
Completed by your machine builder or integrator
A customer-appointed partner designs the metal housing and permanent mount, connects plant electrics and controls, completes safety and conformity work, and commissions the system on site. We provide the specifications and answer integration questions.
05 / Team
Founder · Computer vision and edge systems
Jeremias selects cameras and lighting, prepares image data, trains and evaluates computer vision models, develops data-analysis software and deploys models to edge hardware.
He built RallyTrace’s tracking pipeline, TensorRT runtime and frame-linked review interface. His project work includes industrial image analysis, medical R&D prototypes and sports tracking in startup, company and independent research settings.
Co-founder · Electrical engineering
Sven supports prototype electronics, sensor connections and electrical interfaces between the camera, edge computer and surrounding system.
06 / Contact
We usually start with a short video call. We can look at the task, available samples and operating conditions together, then decide what a useful first test would be. Leave an email address or phone number and we’ll suggest a time.
Prefer email? j.buchendorfer@gmail.com