66 lines
2.6 KiB
Markdown
66 lines
2.6 KiB
Markdown
+++
|
||
title = "Analyzing Truck Load Using Image Recognition"
|
||
date = 2026-02-12
|
||
|
||
[taxonomies]
|
||
categories = ["Software"]
|
||
|
||
[extra]
|
||
author = "Emil Miler"
|
||
+++
|
||
|
||
A friend of mine asked for help expanding an image recognition system he is implementing at work. The [Roboflow](https://roboflow.com/) pipeline analyzes images of open trucks and tags all visible cargo with the appropriate label and dimensions.
|
||
|
||
The goal was to extend the output to display the percentage of the truck’s maximum capacity being used. To achieve this, I integrated a custom Python script into the existing pipeline.
|
||
|
||
<!-- more -->
|
||
|
||
Roboflow is a platform for managing image datasets, training computer vision models, and running image recognition workflows. My friend had already trained a functioning model before I got involved.
|
||
|
||
|
||
|
||
Since the system operates on 2D images, we can calculate only the rough two-dimensional surface coverage of the truck, not the full cargo volume. For their specific use case, however, this level of estimation is sufficient.
|
||
|
||
I extended the existing pipeline by adding a new branch with a custom script block called _Total Coverage_. This block allows for custom logic and can interact directly with the [Roboflow API](https://inference.roboflow.com/workflows/about/) to access workflow data and inference results.
|
||
|
||
|
||
|
||
Below is the Python script I implemented to extract the inference data and calculate both the total surface coverage and the number of detected objects.
|
||
|
||
```python
|
||
def run(self, model_output) -> dict:
|
||
object_count = len(model_output)
|
||
truck_area = 0
|
||
cargo_area = 0
|
||
|
||
for xyxy, mask, confidence, class_id, tracker_id, data in model_output:
|
||
w = data.get("width", 0)
|
||
h = data.get("height", 0)
|
||
label = data.get("class_name", "").lower()
|
||
area = w * h
|
||
|
||
if label == "truck":
|
||
truck_area = area
|
||
object_count -= 1
|
||
else:
|
||
cargo_area += area
|
||
|
||
coverage = round((cargo_area / truck_area * 100), 1) if truck_area > 0 else 0
|
||
|
||
return {
|
||
"total_coverage": coverage,
|
||
"object_count": object_count,
|
||
}
|
||
```
|
||
|
||
The tricky part was handling the `model_output` correctly, otherwise the logic is straightforward. Needless to say, this remains an estimation rather than a precise calculation. Here is an example of the script output in JSON format.
|
||
|
||
```json
|
||
"truck_stats": {
|
||
"total_coverage": 51.8,
|
||
"object_count": 8
|
||
},
|
||
```
|
||
|
||
This was an interesting exercise, as it was my first time working with Roboflow and contributing to an image recognition project that is actively used in production.
|