Meng's Page

Traffic Scenario Builder

Thu, 01 Jan 2026 00:00:00 +0000

Berlin Transit Live Monitor

Thu, 13 Nov 2025 00:00:00 +0000

A modern, responsive web dashboard designed for real-time monitoring of public transport departures in Berlin. This tool provides a comprehensive view of transit movements, combining temporal schedules with spatial tracking. The page works on both desktop and mobile browsers and automatically updates vehicle positions every few seconds.

Figure 1. Dashboard on desktop browser.

This project was specifically designed and optimized to turn any spare smartphone or tablet into a dedicated, real-time smart transit monitor. By opening this dashboard in full-screen on an old device, you can create a permanent public transport timetable for your home, hallway, or office.

Figure 2. Android phone displaying the timetable and real-time map (via the Fully Kiosk Browser app).

Figure 3. Kindle displaying the departure timetable.

---

🔗 Data Source

Text-Mining Demo

Sun, 26 Oct 2025 00:00:00 +0000

Summary

This demo explores the use of _Hugging Face zero-shot text classification models_ to analyze multilingual user comments (e.g., YouTube). Two classification dimensions are focused on:

Sentiment: whether the comment is positive, neutral, or negative
Category: what the comment refers to (functionality, UI, or ads)

Selected Models

Three pre-trained NLI-based transformer models are used for comparison:

Model A: facebook/bart-large-mnli
Model B: joeddav/xlm-roberta-large-xnli
Model C: typeform/distilbert-base-uncased-mnli

Classification

Following are 10 manually labeled comments, spanning multiple languages, sentiment types, and categories. Each model classifies each comment for both sentiment and category, correct results are shown in bold in the table below.

Comment	Sentiment			Category
	A	B	C	A	B	C
The video loads fast and never lags. Love it!	positive	positive	positive	functionality	functionality	functionality
The search function works fine, but nothing special.	neutral	positive	positive	functionality	functionality	user interface
I can’t rewind properly anymore, super annoying.	negative	negative	negative	functionality	functionality	functionality
I like how the app looks now, clean and smooth.	positive	positive	positive	user interface	user interface	functionality
I am ok with the UI, it can be better though.	positive	positive	positive	user interface	user interface	functionality
The layout is messy after the latest update.	negative	negative	negative	user interface	user interface	functionality
Way too many ads these days, it ruins the experience.	negative	negative	negative	ads	ads	ads
Ads are skippable now, so it’s not too bad.	positive	positive	positive	ads	functionality	ads
DE: Zu viele nervige Werbungen, es macht keinen Spaß mehr.	negative	negative	positive	ads	ads	functionality
CN: 功能正常，但感觉有点卡。	positive	positive	positive	user interface	functionality	user interface

Performance

Classification performance was compared across three models, Model A (facebook/bart-large-mnli) achieved higher scores in both the Sentiment and Category dimensions.

Next Steps

App Store and Google Play both provide public access to app reviews. By identifying the corresponding app ID (for example, YouTube’s App Store ID is 544007664), these review data can be extracted automatically. Combined with the comment classification method demonstrated above, this enables real-time categorization of user feedback and helps track emerging user needs and pain points.

Resume

Tue, 21 Oct 2025 00:00:00 +0000

Do cyclists disregard ‘priority-to-the-right’ more often than motorists?

Wed, 18 Sep 2024 00:00:00 +0000

---

Background

Do cyclists disregard traffic rules more often than motorists? Answering this is essential for optimizing traffic safety and efficiency as we transition toward autonomous driving in mixed environments. Grounded in naturalistic observations at urban unsignalized intersections, our study utilizes quasi-experimental methods and regression analysis to quantify compliance behaviors. This research is the result of a seamless collaboration between project managers, sensor technicians, data engineers, and human factors researchers, aiming to provide the critical behavioral data needed for safe human-machine coexistence.

Figure 1. The Collaborative Project Workflow.

Methods

Using stationary cameras, a 12-day naturalistic traffic observation was conducted at an urban T-intersection in Braunschweig, Germany. The video footage underwent spatial calibration to establish a real-world coordinate system. Road users were detected and classified, and their movements were converted into georeferenced trajectories. In 202 interaction cases, a car from the right (Ego, with priority) encountered a car or a bike from the left (Foe, without priority). The study examined how the following factors associated with the violation using logistic regression:

Ego’s turning direction (left vs. right)
Foe’s type (car vs. bike)
Relative arrival time
Foe’s lateral position

Figure 2. Traffic Scenarios: Compliance vs Violation.

Key Findings

Cyclists violated the “priority-to-the-right” rule more frequently than motorists.
Road users with priority were more likely to yield when:
- turning right
- facing a bike
- arriving later at the intersection
- facing a road user close to the lane center

Conclusion

This study underscores the importance of implicit communication in mixed traffic. It provides empirical benchmarks for designing human-like autonomous driving systems, which is supposed to be capable of interpreting and responding to nuanced road-user interactions at unsignalized intersections.

Bringing Emotion Theory into the Cockpit

Sun, 19 Nov 2023 00:00:00 +0000

Background

For nearly a century, the scientific debate on emotion has revolved around two fascinating puzzles. First, the nature of emotion: Are joy and sadness distinct entities, like apples and pears, or merely different shades of the same fruit, like red and green apples? Second, the sequence: Do we run because we feel fear, or do we feel fear because we are running? While the academic community has yet to reach a verdict on these theoretical roots, there is one consensus: to truly understand human emotion, we cannot rely on a single source. It requires multimodal measurement, combining physiological signals, facial expressions, and behavioral data.

This principle is now reshaping the automotive world. From enhancing the passenger’s well-being to solving the complex puzzle of trust in autonomous vehicles, accurate emotion detection is key. Motivated by this, I had the privilege during the early stages of my PhD to independently design and execute a comprehensive multimodal driving simulator experiment, aiming to bridge the gap between abstract theory and real-world application

Figure 1. The Research Workflow.

Hypotheses

Drawing upon the Component Process Model (CPM), the driver’s body responses during fear-inducing scenarios would be modulated by specific cognitive appraisals: Novelty and Power. It’s hypothesized that these cognitive checks manifest through distinct channels:

The Novelty Hypothesis: The appraisal of Novelty (high unexpectedness) will be indicated by immediate responses, specifically an increase in Pupil Diameter, Skin Conductance Level, and the activation of Upper Face Action Units.
The Power Hypothesis: The appraisal of Power (low coping potential/control) will be indicated by a delayed response pattern, specifically an drop in nasal temperature and the activation of Lower Face Action Units.
The Temporal Dynamics Hypothesis: In line with the sequential nature of CPM checks, the facical indicators of Novelty are hypothesized to precede the ones of Power, reflecting the cognitive processing order from event detection to coping evaluation.

Methods

Driving simulator experiments were conducted using Virtual Reality to elicit fear through critical traffic events. Adopting a mixed-methods design, the study integrated objective physiological measurements with subjective assessments. The objective data comprised facial infrared thermography, facial Action Units (AUs), and synchronized peripheral physiological signals. These objective metrics were valided with subjective data collected via the Self-Assessment Manikin (SAM) and the Positive and Negative Affect Schedule (PANAS). This comprehensive approach allowed for the cross-validation of appraisal-driven responses, establishing a robust, multidimensional framework for assessing driver emotions.

Figure 2. Overview of the multi-sensor experimental setup.
The diagram shows the integration of physiological, behavioral, and subjective measurement tools within the simulated driving environment (This image is generated by AI).

Figure 3. Mapping of Facial Action Units (AUs) to basic emotion categories.

Key Findings

The studies empirically validated the Component Process Model (CPM) as a robust framework for assessing driver emotions.
Thermal Imaging as an indicator for “Power”
- Fear (characterized by low power or low coping potential) was significantly associated with a decrease in nasal tip temperature (vasoconstriction), effectively distinguishing it from high-power emotions.
Distinct Facial Signatures for Appraisals
- Novelty: Associated with upper face activation (AUs 1, 2, 4, 5, 7).
- Power: Associated with lower face activation (AUs 15, 20, 25, 26).
- Temporal Dynamics: Indicators of Novelty precede those of Power, reflecting the cognitive processing order from event detection to coping evaluation.
Multimodal Synchronization & Latency
- A decrease in Novelty (habituation) significantly correlated with reductions in Pupil Diameter (PD), Skin Conductance Level (SCL), and AU intensity.
- Response Latency: Revealed distinct peak timings for different modalities, with Pupil Diameter peaking earliest (~3.2s), followed by Skin Conductance (~4.1s) and Facial Expressions (~4.3s), and finally Heart Rate (~5.0s).

Modeling Human-Like Interaction Between Cyclists and Vehicles

Thu, 26 Oct 2023 00:00:00 +0000

Introduction

The study, using real-world data, aims to model the interaction between crossing bicycles and right-turning vehicles, attempting to alter the simulator’s guidance at the tactical level. In ten days of video recordings from an urban intersection, 517 valid cases were collected.

Figure 1. Trajectories of crossing bicycles and right-turning vehicles.

Modeling

A logistic regression model of crossing order was employed with features of bicycles (0) and vehicles (1):

speed (v)
speed difference (dv)
distance to crossing point (d2x)
predicted PET (ppet).

The dataset was split into training (80%) and testing (20%) subsets. Models with different feature combinations were cross-validated.

$$ \log\left(\frac{p}{1-p}\right) = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \ldots + \beta_k x_k $$

Simulation

The selected model was implemented in SUMO: At 20 m before the crossing point, the crossing order was estimated, and the lagging object was forced to slow down. A comparison was performed between real data, SUMO’s default model, and this new model.

In reality, 74% of bicycles crossed before vehicles.
In the default model, vehicles were more conservative - bicycles crossed first in about 90% of cases.
The new model brings this ratio closer to 50%, though still not identical to real data.
In the default model, PET distributions vary strongly by crossing order, whereas in the real-world data and new model, PET distributions are similar (yet differ by ~1 s).

Conclusion

The study of interactions between bicycles and vehicles in real traffic contributes to improving simulator realism.
Although the new model did not perfectly reproduce the crossing-order distribution, it significantly improved the PET distribution, bringing it closer to reality.

Human Performance in Critical Scenarios as a Benchmark for Highly Automated Vehicles

Mon, 19 Jul 2021 00:00:00 +0000

Before highly automated vehicles (HAVs) become part of everyday traffic, their safety must be proven.
The use of human performance as a benchmark represents a promising approach - yet, appropriate methods to quantify and compare human and HAV performance remain rare.

By adapting the method of constant stimuli, a scenario-based approach to quantify the limit of human performance is developed.
This method is applied in a driving simulator study, where participants repeatedly face a cut-in manoeuvre on a highway.

By systematically manipulating the criticality of the manoeuvre - defined in terms of time to collision (TTC) - humans’ collision-avoidance performance can be measured.

Key Method

The limit of human performance is identified using logistic regression.
The regression curve and its inflection point allow direct comparison between human and HAV performance.

Conclusion

The presented approach provides a practical means by which HAVs’ safety performance can be evaluated and proven against human driving capabilities.

How Can Driver Emotions Be Quantified?

Tue, 12 Apr 2016 00:00:00 +0000

How can emotions be quantified?

To describe facial expressions systematically, Paul Ekman and his colleagues developed the Facial Action Coding System (FACS) ¹. This system groups visible facial muscle movements and certain head motions into identifiable units called Action Units (AUs). Each AU represents a specific facial muscle activity that can be observed and coded manually or by software.

The table below shows a selection of AUs along with their descriptions and how accurately they were detected by facial recognition software I used in 2016.

Action Unit	Description	Accuracy	N
1	Raise inner eyebrow	89.7 %	175
2	Raise outer eyebrow	88.9 %	117
4	Lower brows	94.3 %	194
5	Raise upper eyelids	95.1 %	102
6	Raise cheeks	92.7 %	123
7	Tighten eyelids	93.4 %	121
9	Wrinkle nose	100 %	75
10	Raise upper lip	90.5 %	21
12	Pull lip corners (smile)	95.4 %	131
14	Dimpler	67.6 %	37
15	Depress lip corners	89.4 %	94
17	Raise chin	86.6 %	202
18	Pucker lips	88.9 %	9
20	Stretch lips	92.4 %	79
23	Tighten lips	63.3 %	60
24	Press lips together	65.5 %	58
25	Open mouth	76.9 %	324
26	Drop jaw	48 %	50
28	Suck in lips	100 %	1

Although there are some individual differences, it is generally assumed that specific facial expressions are universally associated with certain emotions, across people and cultures. The diagram below illustrates how specific combinations of Action Units (AUs) form the prototypical expressions of basic emotions like fear, joy, sadness, and anger².

How can we read a driver’s face?

The Facial Action Coding System (FACS) offers a systematic way to describe facial muscle movements, and that means we can leverage it to assess driver emotions. But here’s the question:

Which facial muscle movements are linked to frustration during driving?

30 participants were invited to take part in a simulated driving experiment. Each driver faced two contrasting traffic scenarios:

🚗 Smooth traffic
🚗 Traffic jam

To enhance immersion, participants were rewarded for reaching their destination within a set time limit: making delays feel more frustrating and emotionally charged. The image below shows the driving simulator setup and one such frustrating moment, captured inside our virtual traffic jam:

Statistical results revealed that several facial muscle movements (Action Units, AUs) appeared significantly more often during traffic jam than in smooth ones. These frustration-linked AUs include:

AU2: Raise outer eyebrows
AU5: Raise upper eyelids
AU6: Raise cheeks
AU9: Wrinkle nose
AU10: Raise upper lip
AU12: Pull lip corners (smile)
AU14: Dimpler
AU15: Depress lip corners
AU17: Raise chin
AU18: Pucker lips
AU23: Tighten lips
AU28: Suck in lips

What is a frustrated facial expression?

While individual Action Units (AUs) are meaningful, facial expressions are usually combinations of multiple AUs. In fact, the AUs that appeared more frequently during traffic jams may not act alone. They likely combine to form complex expressions. To uncover these hidden combinations, I conducted a clustering analysis on the facial data observed during the jam condition. K-Means Clustering was employed. The core idea is simple: group similar data points based on how close they are in space.

The K-Means process works in following steps:

Randomly choose initial cluster centers.
Assign each observation to the nearest cluster center.
Update the cluster centers based on the new groupings.
This cycle repeats until stable clusters emerge.

Facial expression data were grouped into five clusters using K-Means clustering. Each cluster represents a common combination of Action Units (AUs), potentially reflecting a distinct emotional pattern. The radar chart above shows the average activation level (cluster centers) for each AU across the five clusters.

Cluster 4 was the only one that appeared significantly more often during traffic jams.

Cluster 4 is mainly defined by the co-activation of:

AU 9 – Wrinkle nose
AU 18 – Pucker lips
AU 24 – Press lips together

This particular combination may represent the facial expression of frustration in the driving context.

What does a “frustrated face” look like?

AU9 + AU18 + AU24 =

Ekman, P. Friesen, W. V. & Hager, J. C. (2002). Manual for the facial action coding system. Salt Lake City: A Human Face. ↩︎
Ekman, P. & Friesen, W. V. (1978). Manual for the facial action coding system. Palo Alto: Consulting Psychologists Press. ↩︎