Open-source video analytics tools for Caltrans CCTV camera systems, providing benchmarks to assess computer vision solutions and comprehensive procurement specifications for digitalizing transportation infrastructure statewide.
A complete vehicle detection, tracking, and annotation system for generating ground-truth traffic data from CCTV footage. Uses automated AI detection with a human-in-the-loop correction interface for quality assurance.
PySide6/Qt standalone app for macOS & Windows with a nine-workspace tabbed interface. v2.2.0 adds hybrid GPU packaging — a universal installer that ships CPU PyTorch and auto-fetches the matching CUDA build (cu128 for Blackwell, cu118 for older GPUs) on first run.
Automated detection and tracking runs in background threads. Supports multiple video formats (MP4, MOV, AVI, MKV) with JSON track data, correction layers, and COCO format export.
Comprehensive tools for video processing, annotation, and ground-truth generation.
Automatic vehicle detection using state-of-the-art object detection
Persistent ID assignment across video frames for accurate tracking
Draw, resize, and adjust detection boxes frame by frame
Split incorrectly merged tracks into separate vehicle paths
Combine fragmented detections into single continuous tracks
Full history management with Ctrl+Z / Ctrl+Y support
Rectangle and polygon regions of interest with per-ROI counting
Export corrected annotations in standard COCO format for training
Rigorous evaluation of computer vision technologies for processing Caltrans CCTV video recordings across five critical dimensions.
Detecting, tracking, and classifying vehicles under varying camera resolutions, framerates, and weather conditions.
Assessing capability to detect and track cyclists under different lighting, traffic, and weather conditions.
Evaluating accuracy of pedestrian detection and tracking across crosswalk scenarios.
Assessing real-time processing capabilities under different capture setups and computing devices.
Analyzing input/output data volumes and determining storage, transmission, and cloud requirements.
Cameras that are accessible and ready for data collection without technical barriers.
Cameras with minor delays in retrieving videos or post-processing outputs.
Cameras requiring extended time to obtain video recordings or post-processing data.
Cameras of interest where retrieval of recordings or data is uncertain due to technical challenges.
Collection requirements: A minimum of 12 hours of recordings per CCTV camera category, tagged with location, resolution, framerate, camera vendor, weather conditions, lighting conditions, and traffic conditions (peak vs. off-peak).
Measurable specifications to guide Caltrans on procurement and deployment of computer vision platforms for their CCTV network.
Hardware, software, power supply, ingress protection, CPU/GPU, memory, storage, and OS specifications
Mapping road types to achievable detection accuracy levels and recommended processing approaches
Broadband, 3G/4G/5G, and data transmission speed requirements for video recording and processing