Computer Vision Specialist – Real-Time Object Tracking (C# / ONNX)

Overview We have a real-time computer vision system tracking snooker balls using a custom-trained object detection model deployed via an ONNX pipeline. While the underlying detector works well, we experience frame-by-frame glitches during fast gameplay that we want to resolve entirely through tracking, prediction, and post-processing logic. We are not looking to retrain the detection model. All solutions must be implemented in C# and integrate into our existing runtime loop. Key Challenges to Solve: Tracking Jitter: Eliminate minor pixel-level box vibrations when balls are stationary or rolling slowly. Frame Drops on Impact: Prevent the cue ball or object balls from being temporarily lost for a few frames during high-speed collisions due to overlapping bounding boxes or motion blur. Transient Misclassifications: Use multi-frame state verification to stop a ball from momentarily flipping classes (e.g., a fast-moving red ball reading as brown for a few frames due to camera pixel blending). Expected Technical Approach: Implement a Multi-Object Tracking (MOT) framework (such as a C# implementation of ByteTrack or SORT) utilizing a Kalman Filter for linear velocity state estimation. Utilize an independent Kalman Filter per ball to handle state estimation and bridge temporary tracking gaps using physics-based path predictions. Apply strict track management lifecycle rules to maintain consistent ball IDs and prevent erratic classification flips. Technical Requirements: Proven expertise in Computer Vision, Multi-Object Tracking (MOT), and Kalman Filtering. Strong proficiency in C# (experience with OpenCVSharp or Microsoft.ML.OnnxRuntime is a major plus). Ability to deliver clean, optimized code suitable for real-time inference loop constraints.