How PhasorLab Achieved Accurate Human Depth Detection up to 10 Metres with Mask RCNN and a Single CCTV Camera

Client

PhasorLab is a deep technology company specialising in high-precision network-based positioning. Its core technology — Hyper Sync Net — achieves sub-nanosecond time synchronisation and sub-parts-per-billion frequency synchronisation, enabling centimetre-level target tracking accuracy in positioning systems suitable for both indoor and outdoor deployment. PhasorLab engaged Ajackus to extend its capabilities into computer vision, adding human body detection and depth measurement to complement the positioning infrastructure the company had already built.

Challenge

Building a Custom Depth Model for a Specialised Use Case

Off-the-shelf depth estimation libraries were not designed for PhasorLab’s specific requirements: accurate distance measurement up to 10 metres, operating from a 2D image stream produced by a standard CCTV camera, in real-world environments with variable lighting and occlusion. Developing a model that met these requirements demanded significant research into depth detection algorithms and careful selection from available approaches — with the wrong choice potentially introducing measurement errors that would undermine the centimetre-level accuracy PhasorLab’s positioning technology promised.

Human Body Detection in a Continuous Image Stream

Identifying individual humans within a continuous video stream — rather than in static images — introduced challenges around consistent detection across frames, robustness to partial occlusion, and accurate instance segmentation (distinguishing individual people rather than detecting a group as a single mass). Body shape recognition at regular intervals required a library capable of object segmentation at instance level, not just object classification.

Single CCTV Camera Constraint

PhasorLab’s integration requirements specified that the solution must function with a single standard CCTV camera. This constraint ruled out stereo-vision approaches that rely on multiple cameras for depth calculation and required the depth estimation model to infer 3D information from a 2D image — a more technically demanding problem that required careful algorithm selection.

Goals

The project focused on building a computer vision solution for human detection and depth measurement compatible with PhasorLab’s existing positioning infrastructure.

Journey

Ajackus developed PhasorLab’s computer vision solution across four structured phases, from algorithm research through integration testing.

Phase 1: Algorithm Research and Library Evaluation

The Ajackus team conducted a structured research phase before committing to any specific algorithm or library. For body detection, multiple object detection and segmentation libraries were evaluated against PhasorLab’s requirements — including performance on human-specific shapes, instance segmentation capability, and frame-rate compatibility with continuous video streams. For depth estimation, depth detection algorithms were assessed for their ability to produce accurate distance measurements from monocular (single camera) 2D images in real-world conditions.

Phase 2: Body Detection Module with Mask RCNN

The Ajackus team selected and deployed the Mask RCNN library for the body detection component. Mask RCNN was chosen for its superior instance segmentation capability — the ability to identify and isolate individual humans within an image, rather than simply detecting the presence of people. This precision was essential for PhasorLab’s use case, where distinguishing individual subjects within a scene is required for accurate per-person positioning. The module was implemented to process continuous image streams at regular intervals, providing consistent detection output to the depth measurement layer.

Phase 3: Depth Estimation with DenseDepth

For the distance measurement component, the Ajackus team selected the DenseDepth algorithm model following comparative evaluation of available monocular depth estimation approaches. DenseDepth was chosen for its accuracy in estimating depth from single 2D images — producing distance measurements compatible with PhasorLab’s known area map data. The integration of DenseDepth with the Mask RCNN output allowed the system to calculate the distance between each detected human and the camera, completing the depth measurement pipeline up to the 10-metre requirement.

Phase 4: Integration and Testing

The combined body detection and depth estimation pipeline was integrated into PhasorLab’s existing system architecture and tested across both indoor and outdoor environments. Edge cases — including partial occlusion, multiple simultaneous subjects, and variable lighting conditions — were addressed during testing to ensure reliable performance in real-world deployment conditions.

Results

PhasorLab’s positioning system now incorporates human body detection and depth measurement, extending the practical scope of its centimetre-level positioning technology to real-world human tracking applications.

What went well:

Technical Achievements

• Human body detection implemented using Mask RCNN, providing accurate instance-level segmentation of individual humans within a continuous CCTV image stream
• Depth measurement implemented using the DenseDepth algorithm, enabling accurate distance calculation between detected humans and the camera up to 10 metres from a monocular 2D image input
• Full solution operating from a single standard CCTV camera — no specialised hardware or multi-camera setup required
• System integrated with PhasorLab’s existing network-based positioning pipeline and validated in both indoor and outdoor environments

Business Impact

• PhasorLab’s positioning system now incorporates human body detection and depth measurement, extending the practical scope of its centimetre-level positioning technology to real-world human tracking applications
• The single-camera architecture significantly reduces hardware requirements for deployment, lowering the cost and complexity of installing PhasorLab’s positioning system in new environments
• The solution is compatible with standard CCTV infrastructure that is already widely deployed in commercial and industrial settings, enabling retrofit adoption without new camera hardware