Aerial-Cars

Object dection and speed estimation.

export with id

Training

As I have a relatively low-end computer, I only used yolov8n, which means there is wide room for improvement if I use a more powerful one, such as yolov8l or yolov8x.
28 epoch - 1.3 hours
367 images - 416x416
dataset details can be found here
What I observed in the last few epochs is that the mAP50 and mAP50-95 barely improve. I ran a few more epoch after with (patience=4) and it stops prematurely, which means the model stops improving. However, at the same time, the patience value is really low to draw any meaningful conclusion.

result latest training result (yolov8n).

the model trained mostly on “far view” dataset, such that I tried to use it on this video

Calculations

The idea is fairly simple: since the camera is positioned top-down, if we know the pixel distance a car move in the frame, we can translate it into real-world measurements.
I use the following source:

road_graph

highlight_data

to_real_measure

Based on these, we can see that: speedX = distanceX * meter_per_pixel * fps (each frame)
To get the distance, simply store the id of each car in a dictionary, then compare and update it every frame (or any duration of time).

export

Initial approach (motion detection)

My initial approach is to use Motion detection + Custom tracking based on distance, but later I decided to use yolov8 for expansion potential and learning purposes.

tracking_tracking_in_roi

initial idea of tracking + assigning id in ROI (region of interest)

roi

results of the initial ‘motion detection’ approach