Video and Image AnalyticsVideo and Image Analytics
Scraawl PixL® analytics leverage cutting-edge machine learning techniques---
FACE DETECTION AND TRACKING
PixL uses state-of-the-art deep learning algorithms to detect and recognize faces in unconstrained conditions including non-frontal faces. For each detected face, the module uses a deep neural network to extract a low dimensionality facial feature vector that is discriminative of the face and performs clustering to group similar faces. A tracker is integrated with face detection to correlate continuous occurrences of faces in the video. Tracks of similar faces are grouped into a single entity and analysts can use the analytics drill down pages or the player to navigate through the detections and tracks.
PixL's face recognition analytics allow analysts to recognize individuals of interest by matching faces detected within a video or set of images against a user-provided database of faces. PixL performs off-line analysis using deep learning of the user-provided images to extract user-specific feature vectors and compares them against faces detected in the video. For commercial use cases, PixL comes pre-configured with a celebrity dataset for celebrity recognition. The celebrity dataset includes over half a million images of 50,000 celebrities. PixL provides analysts with the tools needed to update, label, and maintain their own custom image/face datasets against which face recognition can be performed.
OBJECT DETECTION & TRACKING
PixL uses advanced artificial intelligence and machine learning models to detect categories of objects. The object detection modules are combined with a tracker to detect and predict the positions of the objects. The tracker adaptively learns and updates the appearance of an object using a discriminative learning method to distinguishing between the object and the surrounding environment. This results in persistent tracking that is robust to natural image changes and occlusion. Tracks of similar objects are grouped into a single entity and analysts can use the analytics drill down pages or the player to navigate through the detections and tracks.
To address issues related to low resolution data and the overhead camera angle, PixL Satellite Object Detection Analytics leverage specially trained machine learning models to identify objects of interest in satellite imagery. These analytics include a deep learning-based model that was trained using a satellite image dataset containing millions of labelled objects. The model predicts the positions of bounding boxes, the object classes, and their probabilities. A tracker is integrated with satellite object detection to correlate continuous occurrences of objects in the video. Tracks of similar objects are grouped into a single entity and analysts can use the analytics drill down pages or the player to navigate through the detections and tracks.
In addition to a suite of analytics, the Scraawl PixL® dashboard includes a rich set of tools to view videos, explore the analysis, share, and export results. Some examples of these tools and features include:
For intelligence and surveillance applications, PixL's co-occurrence analytics can identify when and where the tracks of entities overlap. For each overlap, or co-occurrence, details of the time of overlap, frame IDs, duration of overlap, number of overlaps, and details of overlapping faces are provided. These analytics provide the capability to identify two or more people who appear in the same scene and other scenes in which the same individuals are seen together. PixL’s interactive user interface allows the analyst to search through co-occurrences and view the frames associated with each co-occurrence, giving context to the instance of each overlapping track.
Enables analysts to view the original and fully analyzed video with bounding boxes and annotations. PixL's player can enhance viewing capabilities with controls for zooming, panning, saturation, and hue.
OPTICAL CHARACTER RECOGNITION (OCR)
Extract and translate text embedded in photos