Smart Video Surveillance: Enhancing AI Features in Collaborative Projects
As part of a partnership engagement, our team contributed to the development of software for a real‑time video processing platform. The focus was on increasing throughput, synchronizing analytics across multiple streams, and ensuring detection overlays are rendered reliably and with minimal latency.
Customer
Insentry — a leading provider of integrated smart video surveillance solutions.
Task
Improve performance and synchronization of video analytics in an already deployed platform. The next development stage required:
-
Higher processing throughput for many concurrent streams;
-
Grouped processing of video feeds and synchronized rendering of neural network outputs;
-
Minimizing visual noise and latency so operators can act in real time, especially for wearable cameras, drones and robotic devices.
Solution
We delivered an integrated solution addressing transport and decoding delays, frame‑rate variability and timing differences across streams. Key elements:
-
Synchronization mechanism for graphical annotations that aligns neural‑network detections across multiple sources;
-
Techniques to reduce buffering and prediction of object trajectories so annotations appear accurately and without distracting jitter;
-
Clustered resource management and autoscaling for neural processing, enabling efficient shared use of compute similar in concept to high‑density inference frameworks.
Technical implementation
-
Annotation synchronization: implemented a pipeline that accounts for network and decoder latencies, source frame rates and timestamp drift to present coherent overlays across many cameras.
-
Low‑latency rendering: built on Media Source Extensions / WebCodecs‑style principles; adapted object‑tracking algorithms to use motion vectors encoded in video frames for trajectory prediction and timely annotation placement.
-
Distributed inference and autoscaling: designed a mechanism for pooled compute resource allocation (functionally comparable to Nvidia Triton) to scale group processing of neural models and optimize resource use across >100 cameras.
-
Architecture choices prioritized browser‑based client rendering to avoid desktop dependencies and support cross‑device access to the monitoring interface.
Result
-
Real‑time multi‑camera analytics: the system reliably overlays detections on many streams with minimal latency and no visual noise.
-
Operator effectiveness: cleaner video and synchronized annotations reduce operator load and improve reaction accuracy.
-
Scalable deployments: autoscaling and shared inference reduce infrastructure costs for large installations (100+ cameras).
-
Flexible scenarios: supports wearable cameras, drones and robotic platforms where low feedback latency is essential.
Why this project is unique
-
Web‑first client: the entire front end is browser‑based, eliminating the need for dedicated desktop clients and enabling access from any device.
-
Robust distributed design: supports replication, failover and load balancing for high availability in large distributed installations.
-
Practical latency reduction: combining motion data extracted from frames with predictive tracking allowed precise annotation timing while lowering compute load.
Technologies
Java 11, Spring Framework, Apache Maven, Hibernate, Cassandra, PostgreSQL, ActiveMQ, Docker