cs.CVSep 30, 2026

CamAgent: An LLM-Agent Framework for Multi-Species Camera-Trap Workflows

Authors: Yutong Deng, Qi Song, Xi Guo, Tianming Wang, Lei Bao, Jianping Ge

Organizations: Faculty of Arts and Sciences, Beijing Normal University, Zhuhai, China. · College of Life Sciences, Beijing Normal University, Beijing, China. · National Forestry and Grassland Administration Key Laboratory for Conservation Ecology of Northeast Tiger and Leopard, Beijing, China. · Faculty of Geographical Science, Beijing Normal University, Beijing, China.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Camera traps accumulated vast, multidimensional data for wildlife monitoring, yet translating raw media archives into meaningful ecological insights remains highly fragmented. Current research workflows require laboriously stitching together disparate analysis tools and scripts, creating steep programming hurdles and complicating end-to-end spatiotemporal analyses. To overcome this fragmentation, we present CamAgent, an autonomous Large Language Model (LLM) agent framework that integrates camera-trap analytical workflows into a unified intelligent ecosystem. CamAgent interprets natural-language ecological intent, schedules computational routing, and executes specialized tools spanning computer-vision perception (e.g., SpeciesNet), CamtrapDP-compatible data management, detection-corrected occupancy modeling, temporal activity analysis, and species co-occurrence networks. The framework automates multi-stage analytical pipelines while maintaining essential data-quality controls and analytical conventions. Consequently, CamAgent significantly reduces manual programming overhead for conservationists, establishing a transparent, scalable, and fully integrated paradigm for camera-trap ecology. Our project is available at https://anonymous.4open.science/r/artifact72c6f4.

Figures & tables

Figure 1: Study area and camera-trap sampling footprint in Northeast China Tiger and Leopard National Park. The inset locates the study area in northeastern China relative to neighboring countries and the Japan Sea. Hatched grid cells show the camera-trap sampling footprint over a digital elevation model spanning 1–1,478 m. The empirical package used for the six natural-language query experiments contained 1,376,017 media records from 17,129 camera identifiers between 1 January 2023 and 1 September 2024.
Common nameScientific nameMedia recordsEvents
Siberian tigerPanthera tigris altaica7,7456,570
Amur leopardPanthera pardus orientalis11,7049,882
Sika deerCervus nippon625,249345,300
Siberian roe deerCapreolus pygargus727,090505,386
Asiatic black bearUrsus thibetanus4,2293,614
Table 1: Media records and independent events for the five species in the bounded empirical package. Events were constructed within each species–camera combination using a 30-min minimum gap (Section • ‣ 2.2.2 ). All retained records span January 2023 to September 2024.
Figure 2: The framework of CamAgent. A user question (left) enters the Agentic Reasoning Core (top), which draws on the Data Layer (middle band) and the Ecological Tool Modules (bottom band) and returns a structured answer, figures, and an exportable report (right). Arrows show the direction of data flow.
Figure 3: The TEMP planning and execution record. Left: the submitted question, the planning trace, the parsed intent, the eight-call plan, and the reflection check. Right: the execution status of each call, the returned activity and overlap results, and the final answer.
NameQuestion typeQuery sampleOutputs
QUALData-quality assessmentAssess the Siberian tiger records for missing fields, invalid timestamps, possible label noise, and duplicate-like records. Explain which downstream ecological analyses remain supported.Required-field completeness, timestamp checks, record-schema flags, duplicate-like groups, analytical restrictions, and Fig. 4 .
TEMPTemporal activity and overlapWhat is the activity pattern of the Siberian tiger, and how much temporal niche overlap does it have with other species?Tiger activity distribution, activity-window shares, pairwise overlap coefficients, uncertainty intervals, and Fig. 5 .
SPATSpatial distribution and hotspotsWhere are Siberian tiger detections concentrated across the camera network, and which cameras are the main detection hotspots? Return a privacy-safe map and do not interpret hotspots as density or occupancy.Camera-level hotspot counts, a privacy-masked grid-cell map, detection-only interpretation limits, and Fig. 6 .
COMMDataset-wide community inferenceUsing the current camera-trap data, perform dataset-wide community inference across all eligible species rather than a hard-coded focal list. Estimate observed and Chao2 richness, Hill alpha and gamma diversity, Jaccard and Bray–Curtis beta diversity, partition Sørensen beta diversity into turnover and nestedness components, construct a species-accumulation curve, and test conditional shared-camera associations with false-discovery-rate correction. Return figures.Dataset-wide diversity estimates, a species-accumulation curve, FDR-corrected shared-camera associations, and Fig. 7 .
MOVECamera-detection adjacency hypothesesFor Amur leopard detections separated by no more than 72 hours and 100 km, summarize candidate inter-camera distances, directions, and repeated camera pairs. Treat them as adjacency hypotheses, not individual tracks or confirmed corridors.Distance and direction summaries, repeated directed camera-pair counts, interpretation limits, and Fig. 8 .
INTEGIntegrated ecological synthesisAnalyze Siberian tiger diel activity and overlap with Sika deer, identify its main detection hotspots, summarize shared-camera occurrence between the two species, check the relevant data quality, and return a concise ecological report with figures and limitations.Tiger activity, tiger–sika deer temporal overlap, shared-camera summaries, hotspot counts, and Fig. 9 .
Table 2: Representative natural-language ecological questions and principal outputs for the six CamAgent experiments.
Figure 4: Data quality for 7,745 Siberian tiger records. Left: completeness of the six required fields. Right: counts of missing fields, invalid timestamps, suspicious timestamps, schema flags, and duplicate-like groups.
Figure 5: Activity overlap between Siberian tiger and four sympatric species. Solid blue curves show tiger activity, dashed colored curves the comparison species, and gray shading the overlap. Sika deer (upper left), roe deer (upper right), black bear (lower left), Amur leopard (lower right). Each panel gives Δ\Delta with its 95% bootstrap interval.
Figure 6: Siberian tiger detections in 0.1-degree grid cells. Numbered orange symbols mark the ten highest-count cells; the bars give event and camera counts for each. Camera identifiers are withheld.
Figure 7: Community summary. Upper left: eligible-species richness by 0.1-degree grid cell. Upper right: species-accumulation curve with observed and Chao2 richness. Lower left: event counts by species. Lower right: standardized effect sizes for shared-camera associations after Benjamini–Hochberg correction.
Figure 8: Amur leopard links within 72 h and 100 km ( n=9,466n=9{,}466 ). Left: link counts across eight directional bins. Upper right: distance distribution with median, mean, and 90th percentile. Lower right: the five most frequent directed camera pairs.
Figure 9: Siberian tiger and sika deer. Upper left: camera sets for each species and their overlap. Upper right: activity overlap (same estimate as Fig. 5 ). Lower left: nearest same-camera cross-species gaps within 24 h. Lower right: the five highest tiger event counts by camera, identifiers withheld.

Explore similar work

CardsList

  1. Counting Animals in Camera-Traps Image Sequences without Count Labels: Winning Solution to the iWildCam 2021 Challenge

    Sep 3, 2026Fagner Cunha, Juan G. Colonna, Eulanda M. dos SantosRoadside CamerasSpecies Identification

    Federal University of Amazonas · Manaus, Amazonas, Brazil

  2. Democratising Camera Trap AI: An Open-Source Model for Detecting UK Mammals

    Jun 9, 2026Paul Fergus, Philip Stephens, Russell A. Hill +8Roadside CamerasBiodiversity Monitoring

    1Liverpool John Moores University, Byrom Street, Liverpool, L3 3AF, UK. · 2Durham University, Stockton Road, Durham, DH1 3LE. · 3MammalWeb, c/o Durham Wildlife Trust, Houghton-le-Spring, DH4 6PU +6

  3. Automated Species Identification in Camera Trap Images for Wildlife Conservation

    Sep 28, 2026Nowshin Amin, Nafisa Tabassum Oyshi, Tahmid Abrar Zidan +2Species IdentificationRoadside Cameras

    Department of Computer Science and Engineering Brac University