Object Detection — Level 8: Video Speed Estimation + Data Export

Approach

Extends Level 7's tracking with per-vehicle speed estimation and a structured data extraction pipeline.

# Pixel displacement between tracked positions / calibrated pixels-per-meter / elapsed time
pixel_dist = ((x_new - x_old)**2 + (y_new - y_old)**2) ** 0.5
meters_per_sec = (pixel_dist / PIXELS_PER_METER) / seconds
mph = meters_per_sec * 2.23694

Explicitly stated as an approximation — true speed requires real camera calibration (e.g., a surveyed lane-width reference), which is stated both in the on-screen video overlay and in terminal output. That honesty is a deliberate part of the narrative: an uncalibrated number presented as precise would undermine credibility the moment someone asks how it was calibrated.

Visual feature: bounding boxes turn red above a configurable mph threshold, green at or below it — recalculated fresh every frame per vehicle, so it toggles live and accurately reflects current speed.

The "Total vehicles tracked" figures below (e.g. 46 in clip 1) are unique track IDs from the same BoT-SORT tracker as Level 7, not a verified physical count — an upper bound, since an occluded vehicle that gets reacquired under a new ID is counted twice. See Level 7 for the full explanation and the production alternative (line-crossing / zone counting).

Real output: clip 1

A sample of the actual per-vehicle rows from this run (47 vehicles tracked total).

ID	Class	Avg MPH	Max MPH	Min MPH	Frames	Over threshold
1	car	14.1	23.2	3.1	74	YES
16	car	15.3	24.3	2.3	74	YES
20	bus	4.8	28.7	0.0	147	YES
26	car	18.2	24.2	12.1	49	YES
45	car	17.2	32.3	6.5	79	YES
48	truck	12.5	22.3	3.2	66	YES
67	car	3.6	8.7	0.0	128	no
71	car	31.3	33.1	29.6	7	YES

Clip 1's speeds look like real highway traffic — mostly single digits to low 30s mph. Clip 4, from a different camera angle, reads wildly implausible speeds for the exact same code and math — one vehicle averaging 157 mph, another peaking past 275 mph:

ID	Class	Avg MPH	Max MPH	Min MPH	Frames	Over threshold
8	car	61.0	121.0	19.5	77	YES
11	car	99.7	206.9	1.8	42	YES
120	car	121.8	163.0	81.4	13	YES
139	car	157.5	193.6	108.6	10	YES
154	car	133.3	155.1	78.7	19	YES
157	car	126.0	275.5	62.3	7	YES
176	car	118.8	243.2	50.3	48	YES

No camera calibration in either case. That gap is the clearest possible demonstration of why an uncalibrated pixel-to-mph conversion is an approximation, not a measurement — clip 4 flags 158 of its 167 tracked vehicles (95%) as "over threshold," versus 0 of 10 in clip 3, the slowest/most static of the four.

↓ speed_data1.csv (clip 1, realistic) ↓ speed_data4.csv (clip 4, uncalibrated)

Demo Clips

Four tracked clips, in order. Clip 1 matches the realistic speed data above; Clip 4 matches the wildly implausible speed data above.

Terminal output

--------------------------------------------------------------------------------
Total vehicles tracked: 46
Vehicles exceeding 10mph threshold: 40
Saved annotated video as output_speed1.mp4
Saved speed data table as speed_data1.csv

Terminal output

--------------------------------------------------------------------------------
Total vehicles tracked: 21
Vehicles exceeding 10mph threshold: 20
Saved annotated video as output_speed2.mp4
Saved speed data table as speed_data2.csv

Terminal output

--------------------------------------------------------------------------------
Total vehicles tracked: 10
Vehicles exceeding 10mph threshold: 0
Saved annotated video as output_speed3.mp4
Saved speed data table as speed_data3.csv

Terminal output

--------------------------------------------------------------------------------
Total vehicles tracked: 167
Vehicles exceeding 10mph threshold: 158
Saved annotated video as output_speed4.mp4
Saved speed data table as speed_data4.csv

Data extraction pipeline

Full per-vehicle speed history collected across the entire video — every reading, every frame, per tracked ID.
End-of-run printed summary table: vehicle ID, class, avg/max/min mph, frames tracked, over-threshold flag.
CSV export — a genuine structured, exportable data artifact, not just a video to watch once.

The video isn't just producing a visual — it's extracting structured, exportable data per object, exactly the kind of output a real manufacturing or logistics deployment would need to feed into a dashboard or reporting system.