Vehicle Classification and Re-ID

The usage of cameras and video data for everyday use has skyrocketed because of a few well-timed advances - faster internet connectivity, more storage with smaller devices, and better camera technology. The corresponding research in analytics for video has similarly boomed.

The SAFR (Small Accurate and Fast Re-ID) project works on an end-to-end traffic monitoring system for real-time pedestrian and vehicle tracking, vehicle recognition, object search, and event monitoring. We are developing a framework to perform these tasks in real-world camera networks, which turn out to be heterogeneous, meaning they encompass a variety of camera types all with different resolutions, scales, and artifacts. Further, as new cameras are added, the camera network composition and heterogenaity also changes.

See our papers at:

  1. ATEAM: Knowledge Integration from Federated Datasets for Vehicle Feature Extraction using Annotation Team of Experts.
  2. Small, Accurate, and Fast Re-ID on the Edge: The SAFR Approach. EDGE 2020
  3. Robust, Extensible, and Fast: Teamed Classifiers for Vehicle Tracking in Multi-Camera Networks. IEEE CogMI 2019

The usual approaches for the traffic monitoring tasks have primarily focused on improving accuracy in homogeneous camera networks. Our work deals with the significant harder heterogeneous camera network problem. Our traffic monitoring system (in-progress) performs the following tasks:

  1. Object Detection: Dense object detection in video streams has long been an integral part of video analytics. Any traffic monitoring requires being able to see the relevant objects, such as vehicles and pedestrians. Our system performs object detection in heterogeneous camera networks by dynamically generating and using high-accuracy real-time models for each camera given its profile (resolution, scale, artifacts, etc).

  2. Attribute Detection: Each vehicle seen in the network needs to be recorded with its global attributes, such as color, vehicle type, and brand, if available. Such attributes can significantly reduce load in searching for vehicles in Amber Alerts or APBs, for example.

  3. Vehicle Re-identification: Re-identification task requires following vehicles across cameras and assigning them to correct identities (i.e. license plate); difficulty arises when the license plate is not visible due to lower resolution of surveillance cameras. One big challenge lies in adversarial conditions where vehicles need to be tracked across cameras with different resolutions and angles. We address this along with the inter-class similarity problem, where two vehicles of differnt license plates look the same (i.e. two different white Toyota Corollas).

  4. Event Detection (future): Automated event detection remains a difficult challenge due to the lack of labeled real-world or synthetic data and absence of frameworks for video-based anomaly detection. We plan to work on real=time event detection once prior tasks can be solved effectively.