Techling builds an appearance-based search system that helps investigators track people and vehicles even when faces are covered or plates are missing. It extracts visual details like clothing, accessories, body shape, vehicle type, color, and unique marks, then turns them into searchable vectors. Users can search by image or by describing attributes, and the system re-links sightings across cameras using time and location hints. A GPU-powered pipeline processes footage, stores embeddings, clips, and metadata, and serves fast results through APIs and a clean UI. The platform runs on Kubernetes with autoscaling, index cleanup, and performance monitoring.
When faces are masked or vehicle plates are missing, forged, or duplicated, investigators lose their primary identifiers. This creates major gaps in tracking, as standard recognition systems cannot connect sightings across locations without dependable facial or plate data.
Lighting, camera angles, motion blur, and distance change how a person or vehicle appears. Clothing shades shift, small details disappear, and unique marks become unclear, making manual and automated tracking difficult across different cameras positioned throughout a city or facility.
Without consistent identity cues, teams rely heavily on human review of long video timelines. This increases the chance of oversight, slows investigations, and demands extensive effort to spot repeating patterns like jackets, backpacks, stickers, dents, or other subtle indicators.
When a target moves across multiple cameras, investigators struggle to maintain continuity using appearance alone. Time gaps, crowded scenes, and similar-looking individuals or vehicles often break the chain, causing missed tracks or incorrect associations in the investigation timeline.
Traditional systems cannot search by descriptive clues such as “red jacket,” “black backpack,” “white sedan,” or “car with roof-rack.” This limits the ability to quickly filter large video archives, leaving investigators without an efficient way to narrow potential matches.
The system records detailed appearance features including clothing colors, patterns, body shape, accessories, vehicle make and model, color, stickers, dents and other signs on the exterior. These features are transformed into structured vectors that allow them to be filtered and re-identified across cameras accurately.
Each detected person and vehicle is generated a general-purpose visual embedding, which will match even with varying lighting, angle, distance, and so on. These embeddings assist investigators to re-discover the same target in time and place enhancing tracking even in cases where identities or plates are unobtainable.
Crime investigators can search either through uploading an image to locate visually similar objects or by specifying the main characteristics, including clothing colors, accessories, or vehicle information. The two modes simplify the exploration of large video sets, allowing quick and adaptable exploration of the huge footage libraries.
Similar sightings on various cameras are connected with the help of timestamps and camera positions, and movement patterns. This serves to keep a target straight, avoiding broken tracks and enabling investigators to have a sequence of movements despite slight differences in appearance across cameras.
Every video is directed to Kafka, where workers with GPUs deal with the extraction of attributes in real-time and generation of embedding. This pipeline will ensure continuous processing, good scaling, and reliability when there are many searches or multiple investigators are searching simultaneously.
Person and vehicle embeddings are stored in separate vector collections, whereas attributes are organized as a query filter. By integrating vector similarity and precise attribute filters, one can achieve accurate, high-speed retrieval that serves a variety of investigative purposes and multifaceted search requirements.
The system provides two dedicated APIs: /search/by-image, which returns ranked visual matches with timestamps and camera IDs, and /search/by-attributes, which filters results using colors, models, stickers, text patches, and time ranges. These APIs support fast, flexible investigative queries.
To make navigation easier, the interface provides attribute filters, a scrubber of the timeline as well as recommended camera paths. It also previews thumbnails and clips in storage, which assist investigators to need not waste time in verifying matches, construct evidence timelines, and move through massive video collections with ease.
Embeddings are stored in a vector database, and thumbnails and clips are stored in the MinIO and metadata in PostgreSQL or Elasticsearch. This division guarantees the optimized retrieval, quick indexing as well as scalable storage needed to sustain long term investigations and uninterrupted data growth.
| Layer / Component | Technology / Tools | Purpose |
|---|---|---|
| Video Ingest Stream | Kafka | Handles continuous video events, metadata ingestion, and workload distribution. |
| Processing Workers | GPU-powered Python services, PyTorch / TensorRT | Perform attribute extraction and generate person/vehicle embeddings in real time. |
| Model Functions | CNNs/Transformers for ReID, attribute classifiers | Extract clothing colors, patterns, accessories, vehicle cues, and unique appearance features. |
| Vector Database | Milvus / Pinecone / Weaviate | Stores person and vehicle embeddings; supports hybrid vector + attribute filtering. |
| Structured Metadata DB | PostgreSQL / Elasticsearch | Stores timestamps, camera IDs, attribute fields, track associations, and search indexes. |
| Object Storage | MinIO | Stores thumbnails, clips, and evidence previews used in UI and investigations. |
| APIs | REST APIs: /search/by-image, /search/by-attributes | Provide similarity search and attribute-based search with ranked results. |
| UI Layer | React / Next.js | Offers attribute filters, timeline scrubbing, path hints, and evidence previews. |
| Orchestration & Deployment | Kubernetes (K8s) + GPU Nodes | Deploys and manages scalable components across clusters. |
At Techling, we specialize in elevating efficiency and
achieving cost savings in the mobility and healthcare
industries through our custom Al and ML software
solutions. We are committed to delivering exceptional results with a 100% satisfaction guarantee and a promise of ontime delivery. Partner with us to leverage the power of Al and ML, and take your business to
new heights