University of Michigan
This research advances knowledge on how to model body-worn camera video for use with state of the art deep learning based computer vision video analytics. It will allow body-worn cameras to be better leveraged operationally in public safety. - July 2019
Principle Investigator: Jason Corso
University of Michigan
Body-worn, or wearable cameras have been increasingly adopted by an estimated 6K-18K public safety organizations (PSO) worldwide. The advantages of such adoption are myriad, including transparency and increase in public trust; officer protection from false claims; forensic analysis; and future increased real-time situational awareness between HQ and the edge.
We believe that the key bridge to realizing these advantages is the analytics involving human activities as observed through the body-worn cameras. Although great progress has been achieved in general third-person-view human activity recognition, the transition of these methods to body-worn cameras is non-trivial. Body-worn camera video is more complex, with a greater range of motion, the activities are often only partially observable, the context is harder to infer, and the availability of annotated data is limited.
The objective of this proposal is to develop a new level of analytical capability in body-worn cameras for public safety. Body-worm camera analytics (BOCA) will analyze human activity from body-worn cameras with minimum human effort for data annotation by leveraging available regularity in the data as well as preexisting labeled data from third-person fixed-camera-view scenarios; it will adapt ideas from transfer learning and multi-task clustering to overcome these challenges.
Challenge 1: Body-worn cameras produce a huge amount of data—unannotated data the annotation of which would require a massive human effort.
Solution: Consequently, BOCA proposes a regime based on transfer learning and multi-task to realize unsupervised and semi-supervised analytics.
Challenge 2: Body-worn camera video has a high range of motion, partial observability and varying context; existing standard feature descriptors and models do not transfer well.
Solution: BOCA will establish body-worn-camera-specific features and it will leverage other information, such as presence/absence of objects, or side-knowledge.
Challenge 3: There is no existing activity recognition dataset in the literature that supports body-worn activity recognition benchmarking, nor with synchronized third-person view data.
Solution: In cooperation with our PSO partner, BOCA will we collect a new large-scale dataset for the proposed project and share with the new Public Safety Innovation Accelerator Program community and beyond.
Benefits and Impact Statements
Social Impact: the proposed research is critical and innovative for the broader public safety community, not just law enforcement. PSOs will be able to efficiently process and store indexed body-worn camera data, search it via semantic queries, and use it in forensic analysis. We will work closely with our PSO partner, Oakland PD, to hone BOCA’s impact on society.
Research Impact: research on activity recognition using body-worn cameras will benefit greatly: BOCA will open up a new research direction for computer vision and multimedia understanding to enrich frontier technologies. As an emerging research field, wearable cameras-based activity recognition has become increasingly important in many social, industrial, and business applications, in which existing techniques are rather limited. The technical solutions, dataset, and high-quality publications produced in the project will foster future research development.