Customized and quantized MobileNetV2 to optimize training/inference for fish-eye camera systems.
This security system focuses on tracking people and detecting potentially threatening objects in surveillance footage, particularly from fish-eye cameras that provide wide coverage but introduce distortion challenges.
We customized MobileNetV2 architecture to handle the unique distortion patterns of fish-eye lenses while maintaining detection accuracy across the entire field of view.
The solution includes sophisticated tracking algorithms to maintain identity consistency as subjects move through the monitored space.
Developed a custom data augmentation pipeline that simulates fish-eye distortion and trained the model specifically on this transformed data.
Applied model quantization reducing bit precision to int8 and optimized the inference pipeline, achieving 30+ FPS on edge devices.
Implemented a two-stage detection approach with a fast initial detector followed by a more precise classifier for potential threats.
"The system's ability to accurately track individuals across our wide-angle cameras while identifying potential threats has significantly improved our security operations."
Michael Wong
Security Director, Metro Transit Authority
