AI Vision Cuts Public Art Installation Energy Use by Nearly 37 Percent

Environmental art installations have quietly become some of the most energy-hungry fixtures in modern cities. Interactive light matrices, kinetic sculptures and multimedia projections now anchor plazas and festivals worldwide, drawing crowds with dazzling displays while consuming electricity around the clock. A new study published in Frontiers in Physics proposes a strikingly different way to power them: teach the installation to watch its audience, understand what people are doing, and allocate energy on demand. The result, tested over 30 days of real load data and a live public festival, cut overall energy consumption by 36.94 percent while actually improving the quality of interaction, raising the interactive quality score from 0.72 to 0.93.

The research, led by Shuqi Meng, Tengyu Jiang and Qi Huang of the Jiangxi Institute of Fashion Technology in Nanchang, China, tackles a stubborn inefficiency in how public installations are controlled. Most existing systems rely on static, rule-based energy allocation: pre-programmed schedules and simple occupancy triggers that cannot perceive what visitors actually intend. The consequence is a double failure. During quiet hours, lights and kinetic components burn power for nobody; during peak crowds, the installation responds sluggishly and delivers a suboptimal experience. The team frames this problem within a cyber-physical-social system, or CPSS, an architecture that explicitly links human behavior in the social space, computation in the cyber space, and power control in the physical space.

The first technical pillar is a self-supervised visual perception model the authors call Env-SSL. Unlike conventional deep learning pipelines that require enormous pixel-level annotated datasets, Env-SSL learns from raw video frames captured by anonymized edge-node cameras without any manual labels. It works through cross-view contrastive learning: two random augmentations of the same frame are passed through twin encoder networks, and the model is trained so that features of the same dimension remain consistent across distortions while different dimensions stay orthogonal. The loss function is built on the batch-normalized cross-correlation matrix of the two views, pushing the diagonal toward one and the off-diagonal elements toward zero. This forces the network to discover the essential geometric structure of the installation and its crowd rather than memorizing scene content.

The performance is remarkable for an unsupervised method. On the team’s EnvDesign-22K validation set, which contains 22,450 structured sketches and real-world images collected during the festival deployment, Env-SSL achieved an angle error of just 3.65 degrees in extracting installation boundaries and an orthogonality score of 0.895 in decoupling crowd features from background features. It outperformed fully supervised benchmarks including ResNet-50 and even a Vision Transformer trained on massive labeled data, while running at 15.6 milliseconds per frame on an NVIDIA RTX 4090 edge server. The team notes that the contrastive loss is inherently robust to fog, snow and sudden illumination changes, because a uniform attenuation of light rescales both augmented views equally and is removed by batch normalization.

Once the perception layer has converted the crowd into numbers, the system quantifies a Social Interaction Intent Index. It combines a spatial density function estimated through two-dimensional Gaussian kernel density estimation, an average dwell-time model driven by the divergence of the visual velocity field, and a saturation constant that mimics the psychological transition from novelty to aesthetic fatigue. The result is a single index that tells the optimizer how strongly people are engaging, where they are clustering, and how long they are staying. This is the bridge that previous research lacked: a mathematical mapping from unstructured social behavior to control variables that directly drive LED channels and servo motors.

The second pillar is the energy consumption model itself. The installation is decomposed into heterogeneous loads: LED lighting channels whose power follows a second-order polynomial in the PWM duty cycle, capturing linear driver conversion loss and quadratic thermal loss; kinetic servo actuators whose instantaneous power couples motor angular velocity with load torque and inertial acceleration losses; and constant multimedia and controller baselines. Environmental corrections are applied in real time: a thermodynamic factor adjusts driver efficiency for temperature, and a brightness response function exponentially suppresses pre-allocated lighting power as natural illuminance rises. A dynamic energy budget, lifted during crowd peaks and lowered during troughs, bounds the whole allocation.

Solving this high-dimensional nonlinear allocation problem is where the study introduces its most distinctive algorithm: Reverse-Advantage Particle Swarm Optimization, or RA-PSO. Each particle represents a candidate energy allocation strategy, and the fitness function weighs normalized energy consumption against the interactive quality index, with a penalty term excluding solutions that breach the energy budget. The innovation lies in the reverse-advantage mechanism. When the swarm detects that the global best has not improved for a preset number of consecutive iterations, it forcibly generates a companion particle in the geometrically opposite region of the solution space, biased toward high-fitness regions by a global-best traction term. The companion is admitted only through greedy selection, guaranteeing the population best never degrades. In convergence tests, standard PSO fell into a local trap around the 30th iteration; RA-PSO triggered its reverse jump at the 35th, escaping the inferior basin and reaching a far lower final fitness.

The experimental evidence spans three settings, none of which is purely numerical simulation. A hardware-in-the-loop physical sand table drove the real controllers and execution units of the installation with 30 days of genuine load and environment logs collected from a smart city square in Nanchang. Over 20 independent runs, RA-PSO achieved an energy consumption reduction rate of 36.94 percent with a standard deviation of just 1.12, significantly outperforming the Slime Mould Algorithm at 25.65 percent, Deep Deterministic Policy Gradient at 23.30 percent, and standard PSO at 17.01 percent, all at p < 0.001 in paired Wilcoxon signed-rank tests. The interactive quality score reached 0.93, and end-to-end latency was compressed to 22.4 milliseconds, 47 percent faster than DDPG. The team deliberately accepts a small latency overhead of 3.9 milliseconds relative to standard PSO, because the reverse-advantage mechanism performs one additional objective evaluation per cycle, a cost outweighed by the much larger energy gain.

Ablation experiments confirm that every module matters. Removing the self-supervised visual module dropped the interaction quality score from 0.93 to 0.78, proving that perception of subtle crowd movements is what enables aesthetically matched responses. Removing the reverse-advantage strategy collapsed the energy reduction rate from 36.94 percent to 17.01 percent, identical to standard PSO, demonstrating that the algorithmic innovation is the core of the system’s ability to escape local optima. Disabling environmental coupling triggered frequent hardware overload protection under extreme temperature and brightness fluctuations, confirming the necessity of the social-physical-environment collaborative model.

The field deployment during the City Summer Night Light and Shadow Interactive Festival provided the most vivid demonstration. Energy heat maps showed the installation concentrating power in two detected interaction hotspots while holding cold zones at low-power standby, with the spatial distribution of energy aligning tightly with actual crowd aggregation. When the crowd migrated rapidly around 21:00, the system completed a smooth migration of the energy center within just three control cycles, avoiding jitter or supply delays. The authors emphasize privacy as a condition of the work: all visual processing runs locally on the edge device, no face recognition is performed, raw frames are deleted within 24 hours, and only aggregated statistics are retained. They acknowledge limitations, including reliance on high-end GPU hardware and statistical approximations of deep interaction psychology, and point toward federated learning, digital twins and lightweight NPU deployment as future directions. For now, the study offers a rigorous blueprint for a new generation of public art that is not only intelligent and responsive but genuinely low-carbon, a paradigm the authors argue could extend to the sustainable digital upgrade of urban infrastructure worldwide.

Subject of Research: Vision-based energy efficiency optimization of environmental art installations under cyber-physical-social systems

Article Title: Vision-based energy efficiency optimization of environmental art installations under cyber-physical-social systems

Article References: Meng, S., Jiang, T., & Huang, Q. (2026). Vision-based energy efficiency optimization of environmental art installations under cyber-physical-social systems. Frontiers in Physics, 14, Article 1942665. https://doi.org/10.3389/fphy.2026.1942665

Image Credits: AI Generated

DOI: 10.3389/fphy.2026.1942665

Keywords: Vision-based, energy, efficiency, optimization, environmental, installations, under, cyber-physical-social, systems, scientific research