Jose Gibson
← All Projects

IBVAS Automatic Passenger Counting

Automatic Passenger Counting POC using head/person detection, tracking, and zone crossing logic for edge deployment.

PythonTensorFlow/Keras-style SSD training pipelineYOLOv8s / Ultralytics tracker pathByteTrack/DeepSORT-style tracking planningYAML zone configuration
⚡ Passenger counting depends on stable detection plus stable i…⚡ Overhead head detection and non-overhead pedestrian detectio…⚡ Counting accuracy targets need careful evaluation beyond raw…⚡ Edge deployment constraints require INT8 export and real FPS…

Outcomes

  • ●APC project has both research/training path and practical YOLO tracker path.
  • ●Zone-counting artifact exists with video output and preview image.
  • ●Clear deployment target and metrics were defined for a demo.

Portfolio Highlights

  • →Developed an Automatic Passenger Counting POC with head-detector training, YOLOv8 tracking, and configurable zone-crossing logic.
  • →Designed detector decision gates around AP50, tracking stability, and INT8 edge deployment readiness.
  • →Built zone preview and tracker output tooling to debug passenger flow counting visually.

Snapshot

  • Period: March 2026
  • Source: `aspire7:D:\ML\ibvas`
  • Domain: Automatic Passenger Counting, object tracking, edge AI
  • Target: NXP i.MX93, INT8 TFLite
  • Status: POC/research with YOLO tracker and training pipeline

Portfolio Summary

The IBVAS/APC repo is the missing passenger-counting work. The roadmap targets a demoable Automatic Passenger Counting POC by March 28/29, 2026, with NXP i.MX93 deployment, INT8 TFLite, >15 FPS, and >92% counting accuracy as stated goals. The work includes custom SSD head-detector training, dataset/evaluation planning, YOLO fallback paths, and a YOLOv8s tracker with zone counting.

Stack

  • Python
  • TensorFlow/Keras-style SSD training pipeline
  • YOLOv8s / Ultralytics tracker path
  • ByteTrack/DeepSORT-style tracking planning
  • YAML zone configuration
  • OpenCV/video output
  • NXP i.MX93 and INT8 TFLite deployment planning

What I Built

  • APC roadmap and decision gates for detector quality vs fallback paths.
  • v6/v9 head-detector training pipelines with augmentation, finetune, and evaluation logs.
  • YOLOv8s-based tracker with zone counting.
  • Zone preview tooling and `zones.yaml`.
  • Repository split into `overhead/` and `yolov8n_tracker/` paths.

Key Decisions

  • Treated detector quality as a decision gate: if v9 SSD head detector met AP targets, continue; otherwise pivot to YOLO.
  • Planned DeepSORT/ByteTrack tracking as a required step regardless of detector path.
  • Built YOLOv8s tracker path to get a practical zone-counting POC moving.
  • Used explicit zone configs/previews so counting logic could be inspected visually.

Development Timeline

  1. March 2026: Added v6/v9 head detector training, model checkpoints, evaluation logs, and Colab workflow.
  2. Roadmap captured POC target, i.MX93 target, detector decision gate, and deployment constraints.
  3. Added YOLOv8s-based APC tracker with zone counting.
  4. Restructured repo into `overhead/` and `yolov8n_tracker/`.