Master's Thesis: Function-Aware Target Confidence Estimation for Fused ADAS Sensors

Master's Thesis: Function-Aware Target Confidence Estimation for Fused ADAS Sensors

Location

Göteborg, SE, 417 15

Position Type:

Student

Permanent/Temporary:

Temporary

Transport is at the core of modern society. Imagine using your expertise to shape sustainable transport and infrastructure solutions for the future. If you seek to make a difference on a global scale, working with next-gen technologies and the sharpest collaborative teams, then we could be a perfect match.

Background

Advanced Driver Assistance Systems (ADAS) rely on sensor fusion to detect and track surrounding traffic participants. A fused system may combine camera, radar, and other sensor inputs, but the reliability of a target estimate depends on more than the confidence reported by any individual sensor. Sensor errors can be correlated, and performance can vary with distance, weather, occlusion, road scene, object type, and other operating conditions. Different vehicle functions also have different needs. Adaptive Cruise Control (ACC) should provide useful, dependable operation across its intended operating domain. Emergency braking, by contrast, must avoid unnecessary interventions while still responding to genuine collision threats. A single generic target-confidence score may therefore be unsuitable for both functions. Real-traffic logs and reference sensor data offer an opportunity to study how target confidence can be estimated and calibrated under practical conditions. Alongside conventional probabilistic methods, AI-based approaches can be evaluated in two forms: higher-capacity models for offline analysis, and smaller models with potential to run within the computational constraints of an ADAS system.

Description of the Thesis Work

The thesis will investigate how to estimate the confidence of fused traffic targets using reference data collected in real traffic. The work will distinguish the quality of the target estimate from the decisions made by individual ADAS functions. The study may consider confidence in several aspects of a target, such as:

• Whether a real object exists at the estimated location
• The probability of the object’s class
• The accuracy and uncertainty of position and velocity
• Track stability over time and the likelihood of a false or lost track

The student will develop and compare a transparent probabilistic baseline with suitable AI-based methods. The analysis will account for the fact that sensor errors may share common causes, rather than assuming that camera and radar false detections are independent. The resulting confidence estimates will be assessed in the context of ACC and emergency braking. The thesis will investigate how each function’s requirements affect the interpretation and use of confidence. It will not assume that one confidence threshold is appropriate for both functions, nor that a confidence score alone is sufficient to establish system safety.

Potential Research Questions

• How well can target existence and target-state confidence be estimated from real-traffic reference data?
• How do correlated false detections and operating conditions affect fused confidence estimates?
• How does a calibrated confidence estimate support different objectives for ACC and emergency braking?
• How do probabilistic and AI-based methods compare in calibration, robustness, and data efficiency?
• What accuracy and runtime trade-offs are achievable with a compact model suitable for constrained ADAS hardware?

Potential Areas of Work

• Preparing and aligning logged sensor data with reference measurements and ground-truth labels
• Defining target-confidence outputs and their intended interpretation
• Estimating single-sensor and joint sensor error characteristics
• Developing a probabilistic baseline and comparing it with AI-based methods
• Calibrating confidence estimates and evaluating generalization to unseen drives or conditions
• Exploring model simplification or compression for a real-time implementation
• Assessing function-relevant performance, including ACC availability and emergency-braking false and missed intervention indicators
• Reporting limitations in rare-event estimation and reference-data coverage

Methodology and Evaluation

The work may compare a conventional probabilistic approach with selected machine-learning methods. Depending on data and scope, the AI methods could include a higher-capacity offline model and a smaller model intended for later real-time evaluation. Both should be assessed for probability calibration, not only their ability to rank detections. Data should be divided by drive, route, or recording session where possible, so that closely related frames do not appear in both training and test sets. Evaluation should be broken down by relevant conditions, such as range, object class, visibility, occlusion, and sensor combination. Potential evaluation measures include:

• Confidence quality: reliability plots, Brier score, and log loss
• Target performance: false and missed detections, tracking quality, and state-estimation error
• ACC relevance: availability within the defined operating domain, alongside target-estimation performance
• Emergency-braking relevance: false and missed or late intervention indicators for defined threat scenarios
• Deployment feasibility: inference time and model size, with other resource measures where available

Rare-event rates should be reported with appropriate uncertainty and exposure definitions. For example, a false-intervention rate per driving hour is not interchangeable with a missed-intervention rate per relevant collision-threat encounter. A limited number of observed events should not be presented as proof of a very low failure rate.

Expected Outcomes

The thesis is expected to deliver:

• A documented method for estimating target confidence from reference data
• A comparison of probabilistic and AI-based approaches
• An analysis of sensor-error dependence and its impact on fusion
• An evaluation of confidence calibration across operating conditions
• A discussion of how confidence can be used differently by ACC and emergency braking
• Recommendations for further validation and possible real-time implementation

Suitable Background for the Student

We are looking for a master’s student with an interest in sensor fusion, ADAS, and data-driven system evaluation. Relevant academic backgrounds include:

• Machine Learning and Data Science
• Robotics and Sensor Fusion
• Signal Processing
• Computer Vision
• Automotive or Control Engineering

Experience with Python, probabilistic modelling, machine learning, or automotive sensor data is beneficial. Experience with real-time or embedded machine learning is useful for the model-deployment extension.

Practical Information

Thesis level: Master’s thesis Language: English Starting date:Jan 2027 Application deadline: 30th Nov The final scope wil