For this thesis project, there are 5 different ideas. Read through them all and see which ones interest you. Final idea is decided together with the supervisor/manager.

Category

Electrical or Computer Engineering, Engineering Physics, Mathematics and similar

Scope

Two Master Students, 30hp each

Idea 1: Future aware rate control

Relevant courses

Efficient C (or similar), Automatic Control, possibly Algorithms and Data Structures (mostly for the algorithm aspect)

Background

Standard real time system bitrate control algorithms are reactive, using the size of previously encoded frames to control the video encoding. While latency is always bad, this can be particularly problematic as the algorithm needs to be reactive enough for scenarios where there are large, sudden changes, e.g. a train passing by, but also not be too reactive to create self-oscillations.

There are however other indicative data available in our camera video pipeline that can estimate the amount of current motion in the scene and that is available several frames ahead of it's encoding which could be used to improve the bitrate control algorithm.

Goal

Evaluate the benefits of using future motion indicative data in video bitrate control algorithms.

As there are several options for the data types and how far into the future we can see, a good side goal would be to evaluate them.

Idea 2: Capability based objective video quality

Relevant courses

C/Python (good to know), Mathematical Statistics

Background

Standard objective video quality metrics (e.g. PSNR, SSIM, VMAF) all have issues in determining the surveillance quality of encoded video.

A novel approach would to define new metrics based on the ability of different detectors to detect, classify, recognize and identify different object types. This would mean using e.g. the same analytics application on the raw video and various encoded versions of the video (where information and thus detection capabilities are lost). At far distances this will impact the ability to detect objects ("there is something"), at long distances the ability to classify objects ("it's a human"), at medium distances the ability to recognize objects ("it's the same human") and at short distances the ability to identify the object ("it's Mr. Smith"). In practice, a suite of different analytics applications would likely be the best.

The metric needs to be tuned to handle different light conditions and object distances and false detections might need to be handled somehow.

Goal

Create a capability based objective video quality metric and investigate if this is suitable for measuring surveillance video quality.

Idea 3: Weather aware rate control

Relevant courses

Mathematical Statistics , Efficient C, Algorithms and Data Structures

Background

The local weather conditions can have a large impact on the bitrate generated by the camera and thus our long term bitrate control algorithms could greatly benefit from knowing the current and future local weather conditions when controlling the video encoding. A typical use case would be a 30 day storage on an on camera SD-card, where we'd like to be able to prepare in advance for a coming storm, but also not end up in a scenario where a couple of stormy hours has filled up most of the storage for the full month, severely limiting the video quality for the rest of the time.

The weather estimation could be done either by an on camera neural network or pushed to the camera by a connected web service. This doesn't need to be part of the thesis.

Goal

Investigate if using weather data for the rate controller would be useful for long time storage control.

Idea 4: Temporal difference based objective video quality

Relevant courses

Mathematical Statistics, C/Python

Background

Standard objective video quality metrics (e.g. PSNR, SSIM, VMAF) all have issues in determining the surveillance quality of encoded video.

A novel approach would to define a new…