Extraction and retrieval of vehicle semantics for long-term car park videos
The use of video data as a means of surveillance is no longer a new idea. It is widely known that video data could potentially provide invaluable information for the purpose of analytics, surveillance or security applications. At this very moment, massive amounts of video footages are being recorded...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2020
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The use of video data as a means of surveillance is no longer a new idea. It is widely known that video data could potentially provide invaluable information for the purpose of analytics, surveillance or security applications. At this very moment, massive amounts of video footages are being recorded over the world. While there has been various works on the extraction of human behaviour in a general surveillance setting especially in recent years, not much work are focused on the car park surveillance scenario. The abundance of video data calls for an efficient and effective way of dissecting these raw data while extracting useful information which are easily interpretable. Also, an equally important characteristic for systems is the ease and efficiency of fetching these information when required. The contribution of this work is twofold. First, a framework for the extraction of colour, motion, timestamp, and size information from the video is proposed. The proposed method employs an algorithm that averages out the dominant colour over the course of tracking it and then proceeds to rank it against different hues. As for the motion of the vehicles, the proposed method compiles the relative position/location of each vehicle into trajectory sets. The timestamp information, the object size along with its type are also extracted. A spatio-temporal cube design is adopted to uniquely identify each video footage in an efficient manner such that the stored extracted information can be easily retrieved. Next, the second contribution of this thesis is the construction of a retrieval engine which is able to retrieve video shots based on the given semantics which were extracted. This work proposed an unconventional yet intuitive trajectory input query in the form of a user-described trajectory drawn on the search canvas. The other semantics i.e. colour and timestamp information, are used in the form of keyword-based inputs for the proposed method to locate and rank video shots based on its similarity. The proposed method is then tested using semantics extracted from a month’s worth of data from the surveyed car park using the average Precision@K and the normalised Discounted Cumulative Gain (nDCG) metrics. The proposed method achieved the best precision score of 86% for trajectory retrieval and 91% for vehicle colour retrieval. The average normalised Discounted Cumulative Gain (nDCG) results scored around the region of ∼83% for both trajectory and colour results. This thesis highlights an overall framework for extracting semantics and translating existing raw video data into intuitive representations for efficient querying and retrieval. Contrary to traditional approaches which are labour-intensive and timeconsuming, the proposed method effectively saves time, reduces the cost of manual extraction and video shot retrieval, and is robust under various test scenarios. Various concepts and propositions in this thesis can also be can also be extended to datasets of similar nature. |
|---|