Week 6: July 07 ~ July 13

Preliminaries

This week, I focused on understanding and analyzing the dataset collected last week (week-05) using the follow-line The dataset includes images paired with corresponding linear (v) and angular (w) velocity values, which are essential for training deep learning models for autonomous control. I collected 84,969 data entries, each consisting of an image paired with corresponding action commands—specifically, linear velocity (v) and angular velocity (w). It helps a better understanding of the structure and distribution of the dataset and identifies imbalanced categories that may affect model performance.

To better understand the distribution of control commands and support classification-based training approaches, I began by analyzing and categorizing both linear (V) and angular (W) velocities into discrete classes. I also created visualizations to identify patterns, detect imbalance, and guide potential resampling strategies.

Objectives

• Understand the Distribution of Control Commands
• Categorize Velocities
• Visualize the Dataset
• Identify Imbalanced Categories

Execution

Categorization of Angular Velocity (w)

I categorized the angular velocity (w) values into five discrete classes based on the turning direction and intensity. This approach helps the model focus on steering behaviors rather than predicting exact floating-point values.

Here are the five categories:

• Sharp Left
• Left
• Straight
• Sharp Right
• Right

Angular Velocity (w) Descriptions

Statistic	Value
Count	84,969
Mean	-0.029841
Standard Deviation (std)	0.376109
Minimum (min)	-3.5
Maximum (max)	3
25th Percentile	-0.12933
Median (50%)	-0.00237
75th Percentile	0.09748

We can see the maximum and minimum values of the angular velocity are -3.5 and 3.0. We divided the range into five categories.

Angular Velocity (w) Categories

Category	Range (W)	Description
Sharp Right	W ≤ -1.0	Hard turn right
Right	-1.0 < W ≤ -0.1	Slight right
Straight	-0.1 < W < 0.1	No turn
Left	0.1 ≤ W < 1.0	Slight left
Sharp Left	W ≥ 1.0	Hard turn left

Angular velocity (W): Data distribution among the categories

Image Samples per categories

Categorization of Linear Velocity (v)

I categorized the linear velocity (v) values into four(4) discrete classes based on speed and direction. This approach helps the model focus on speed behaviors rather than predicting exact floating-point values.

The linear velocity values in the dataset range from 5 to 11. I divided them into four categories, each with a range of 1.5 units. (Calculated as: (11 - 5) / 4 = 1.5)

Here are the four(4) categories:

• Slow
• Moderate
• Fast
• Very Fast

Linear Velocity (v) Descriptions

Statistic	Value
Count	84,969
Mean	8.122210
Standard Deviation (std)	1.954871
Minimum (min)	5.000000
Maximum (max)	11.000000
25th Percentile	6.440000
Median (50%)	8.648000
75th Percentile	9.752000

Linear Velocity (v) Categories

Category	Description	Velocity Range (v)
slow	Lowest speed	v ≤ 6.5
moderate	Moderate speed	6.5 < v ≤ 8.0
fast	Fast speed	8.0 < v ≤ 9.5
very_fast	Very fast	v > 9.5

Linear velocity (v): Data distribution among the categories

Frequency of linear velocity (v) vs angular velocity (w)

The analysis indicates how often combinations of linear and angular velocities occur.

References

[1] Follow Line Exercise