evo-pedestrian-counter

Pedestrian Counter for Surveillance Videos

Problem

Given a directory of surveillance-camera video files, count the number of unique pedestrians (people travelling on foot) in each video and write the per-video counts to a structured Excel workbook.

Approach

1. Detect persons with YOLO (COCO class 0).
2. Track across frames so each physical person gets one ID.
3. Filter noise tracks shorter than a visibility window derived from the video's own frame rate.
4. Exclude cyclists via person-bicycle bounding-box overlap.
5. Write to Excel via openpyxl.

Scripts

File	Key function
scripts/utils.py	get_video_files, count_pedestrians_in_video, write_results_to_excel
scripts/run_pipeline.py	run_pipeline(video_dir, output_path, sheet_name, col_filename, col_count)
scripts/validate.py	validate_output(output_path, video_dir, sheet_name, col_filename, col_count)

Runnable Fresh-Agent Example

Adapt the five variables below to the current task instruction.

import sys
sys.path.insert(0, '/app/environment/skills/evo-pedestrian-counter/scripts')

from run_pipeline import run_pipeline
from validate import validate_output

# ---- set from the current task instruction ----
video_dir   = '/path/to/input/videos'
output_path = '/path/to/output.xlsx'
sheet_name  = 'data'
col_file    = 'source'
col_count   = 'total'
# -----------------------------------------------

results = run_pipeline(video_dir, output_path,
                       sheet_name=sheet_name,
                       col_filename=col_file,
                       col_count=col_count)

ok, issues = validate_output(output_path, video_dir,
                             sheet_name=sheet_name,
                             col_filename=col_file,
                             col_count=col_count)
if not ok:
    raise RuntimeError(issues)
print('Done:', results)

How It Works

• A fresh YOLO model is created per video to reset tracker state.
• Sampling interval: derived as max(1, round(fps / 10)) so that roughly ten frames per second are processed regardless of source fps.
• Minimum-detection threshold: derived as max(2, round(fps * 0.5 / sample_every)) so a person must be observed for at least half a second.
• Cyclist exclusion: person tracks that spatially overlap a bicycle track in the majority of co-occurring frames are removed.
• COCO class indices 0 (person) and 1 (bicycle) are public taxonomy.