Background Motion Research Articles

Exploring Techniques for Abnormal Event Detection in Video Surveillance

This study explores and evaluates the effectiveness of various abnormal event detection techniques in video surveillance, addressing challenges such as intrusions, accidents, and suspicious activities. Through a systematic review of related papers, the study reveals the prevalence of traditional methods like background subtraction and motion detection despite their limitations in complex scenarios. It highlights the increasing use of deep learning techniques, particularly CNNs and RNNs, which show promise but require substantial labeled data. The findings underscore the importance of selecting proper detection techniques based on specific surveillance scenarios and emphasize the need for extensive labeled datasets for deep learning methods. The originality of this study lies in its comprehensive review and comparison of various abnormal event detection techniques, providing valuable insights and practical implications for advancing video surveillance systems.

Robust localization of poorly visible tumor in fiducial free stereotactic body radiation therapy

Background and purposeEffective respiratory motion management reduces healthy tissue toxicity and ensures sufficient dose delivery to lung cancer cells in pulmonary stereotactic body radiation therapy (SBRT) with high fractional doses. An articulated robotic arm paired with an X-ray imaging system is designed for real-time motion-tracking (RTMT) dose delivery. However, small tumors (<15 mm) or tumors at challenging locations may not be visible in the X-ray images, disqualifying patients with such tumors from RTMT dose delivery unless fiducials are implanted via an invasive procedure. To track these practically invisible lung tumors in SBRT, we hereby develop a deep learning-enabled template-free tracking framework, SAFE Track. MethodsSAFE Track is a fully supervised framework that trains a generalizable prior for template-free target localization. Two sub-stages are incorporated in SAFE Track, including the initial pretraining on two large-scale medical image datasets (DeepLesion and Node21) followed by fine-tuning on our in-house dataset. A two-stage detector, Faster R-CNN, with a backbone of ResNet50, was selected as our detection network. 94 patients (415 fractions; 40,348 total frames) with low tumor visibility who thus had implanted fiducials were included. The cohort is categorized by the longest dimension of the tumor (<10 mm, 10–15 mm and > 15 mm). The patients were split into training (n = 66) and testing (n = 28) sets. We simulated fiducial-free tumors by removing the fiducials from the X-ray images. We classified the patients into two groups – fiducial implanted inside tumors and implanted outside tumors. To ensure the rigor of our experiment design, we only conducted fiducial removal simulation in training patients and utilized patients with fiducial implanted outside of the tumors for testing. Commercial Xsight Lung Tracking (XLT) and a Deep Match were included for comparison. ResultsSAFE Track achieves promising outcomes to as accurate as 1.23±1.32 mm 3D distance in testing patients with tumor size > 15 mm where Deep Match is at 4.75±1.67 mm and XLT is at 12.23±4.58 mm 3D distance. Even for the most challenging tumor size (<10 mm), SAFE Track maintains its robustness at 1.82 plus or minus 1.67 mm 3D distance, where Deep Match is at 5.32 plus or minus 2.32 mm, and XLT is at 24.83±12.95 mm 3D distance. Moreover, SAFE Track can detect some considerably challenging cases where the tumor is almost invisible or overlapped with dense anatomies. ConclusionSAFE Track is a robust, clinically compatible, fiducial-free, and template-free tracking framework that is applicable to patients with small tumors or tumors obscured by overlapped anatomies in SBRT.

Adaptive Frame Sampling and Feature Alignment for Multi-Frame Infrared Small Target Detection

In recent years, infrared images have attracted widespread attention, due to their extensive application in low-visibility search and rescue, forest fire monitoring, ground target monitoring, and other fields. Infrared small target detection technology plays a vital role in these applications. Although there has been significant research over the years, accurately detecting infrared small targets in complex backgrounds remains a significant challenge. Multi-frame detection methods can significantly improve detection performance in these cases. However, current multi-frame methods face difficulties in balancing the number of input frames and detection speed, and cannot effectively handle the background motion caused by movement of the infrared camera. To address these issues, we propose an adaptive frame sampling method and a detection network aligned at the feature level. Our adaptive frame sampling method uses mutual information to measure motion changes between adjacent frames, construct a motion distribution, and sample frames with uniform motion based on the averaged motion distribution. Our detection network handles background motion by predicting a homography flow matrix that aligns features at the feature level. Extensive evaluation of all components showed that the proposed method can more effectively perform multi-frame infrared small target detection.

RipViz: Finding Rip Currents by Learning Pathline Behavior.

We present a hybrid machine learning and flow analysis feature detection method, RipViz, to extract rip currents from stationary videos. Rip currents are dangerous strong currents that can drag beachgoers out to sea. Most people are either unaware of them or do not know what they look like. In some instances, even trained personnel such as lifeguards have difficulty identifying them. RipViz produces a simple, easy to understand visualization of rip location overlaid on the source video. With RipViz, we first obtain an unsteady 2D vector field from the stationary video using optical flow. Movement at each pixel is analyzed over time. At each seed point, sequences of short pathlines, rather a single long pathline, are traced across the frames of the video to better capture the quasi-periodic flow behavior of wave activity. Because of the motion on the beach, the surf zone, and the surrounding areas, these pathlines may still appear very cluttered and incomprehensible. Furthermore, lay audiences are not familiar with pathlines and may not know how to interpret them. To address this, we treat rip currents as a flow anomaly in an otherwise normal flow. To learn about the normal flow behavior, we train an LSTM autoencoder with pathline sequences from normal ocean, foreground, and background movements. During test time, we use the trained LSTM autoencoder to detect anomalous pathlines (i.e., those in the rip zone). The origination points of such anomalous pathlines, over the course of the video, are then presented as points within the rip zone. RipViz is fully automated and does not require user input. Feedback from domain expert suggests that RipViz has the potential for wider use.

Researches and measurements of high-sensitive seismometers at experimental base «Obninsk»

One of the main directions of new seismometers development for seismological monitoring systems is decreasing the instrumental noise level and nonlinearity distortions of the instrumental seismic channels. In accordance with modern requirements for high-sensitive seismological stations, the instrumental noise of seismometers should be at least three times less than local seismic background noise. By theoretical estimations modern high-sensitive seismometers have high level features: instrumental noise is lower than the minimal seismic background noise, nonlinearity factor is about 0,01% and less. However, the problem of experimental estimation of these parameters in conditions of ambient seismic backgrounds remains to be under consideration. This problem is especially important during development and production of new seismometers. The article presents some results of studying and measuring the instrumental noise level and nonlinearity factor in channels of broadband and short-period seismometers carried out at the experimental base “Obninsk”. To provide a high precision of measurements, the narrow-band digital filtration and spectrum-correlation data procession technique decreasing influence of the seismic background motion on the results uncertainty were used. Due to stable temperature and insulation in the gallery of 30-meter depth, a 20-30% uncertainty for instrumental noise characteristic measuring, which is 10-30 times lower than local seismic background noise, and 0.0004% uncertainty of nonlinearity factor measuring were achieved. The seismological and environmental conditions at the experimental base “Obninsk” of the Geophysical Survey of RAS provide necessary abilities for researche and measurement of instrumental noise characteristics and non-linear distortion factor of high-sensitive seismometers with satisfactory accuracy.

Explicitly encoding the cyclic nature of breathing signal allows for accurate breathing motion prediction in radiotherapy with minimal training data

Background and purposeActive breathing motion management in radiotherapy consists of motion monitoring, quantification and mitigation. It is impacted by associated latencies of a few 100 ms. Artificial neural networks can successfully predict breathing motion and eliminate latencies. However, they require usually a large dataset for training. The objective of this work was to demonstrate that explicitly encoding the cyclic nature of the breathing signal into the training data enables significant reduction of training datasets which can be obtained from healthy volunteers. Material and methodsSeventy surface scanner breathing signals from 25 healthy volunteers in anterior-posterior direction were used for training and validation (ratio 4:1) of long short-term memory models. The model performance was compared to a model using decomposition into phase, amplitude and a time-dependent baseline. Testing of the models was performed on 55 independent breathing signals in anterior-posterior direction from surface scanner (35 lung, 20 liver) of 30 patients with a mean breathing amplitude of (5.9 ± 6.7)mm. ResultsUsing the decomposed breathing signal allowed for a reduction of the absolute root-mean square error (RMSE) from 0.34 mm to 0.12 mm during validation. Testing using patient data yielded an average absolute RMSE of the breathing signal of (0.16 ± 0.11)mm with a prediction horizon of 500 ms. ConclusionIt was demonstrated that a motion prediction model can be trained with less than 100 datasets of healthy volunteers if breathing cycle parameters are considered. Applied to 55 patients, the model predicted breathing motion with a high accuracy.

Robust and efficient FISTA-based method for moving object detection under background movements

Moving object detection is a fundamental task in many video processing applications, such as video surveillance. The robustness and efficiency of background subtraction make it one of the most common methods for detecting moving objects from a video stream. However, adapting background models with moving cameras remains challenging. Major issues include maintaining the model given viewpoint alterations and compensating motions considering depth variations. Moreover, gradual illumination changes, dynamic backgrounds, and complex motions intensify over time in moving camera scenarios, further complicating background model maintenance. In this context, this paper proposes a novel Robust and Online Tensor-based model named ROTAB that incorporates a more implicit consideration of the relationship between sequential frames than the previous methods, allowing for better adaptation to background changes. Moreover, we propose an Improved version of FISTA named IFISTA that employs two strategies to reduce oscillatory behavior and minimize iterations, improving stability and efficiency. Practically, the combination of IFISTA and ROTAB (IFISTA-ROTAB) demonstrates suitable performance for real-time applications. Quantitative and qualitative experiments are conducted on three large-scale datasets, namely CDnet 2014, BMC 2012 and LASIESTA showing the superiority of IFISTA-ROTAB with a gain from two up to seven percent in average.

TOI-6883.01: A Single-transit Planet Candidate Detected from TESS

A new candidate exoplanet, proposed in ExoFOP by authors, it was promoted from TIC 393818343 to TOI-6883.01 at coordinates R.A.(J2000)20:41:10.01 decl.(J2000) + 3:38:17.87 in the Delphinus constellation and its distance is (93.73 ± 0.35) pc from Earth. The target star is a Sun-type having G0 class according to Skiff spectral classification and it has photometric magnitude V(Johnson) = 9.5 mag. We analyzed the transit light curve using Transit Exoplanet Survey Satellite’s ExoMAST tool, and evaluated the Lightcurve Analysis Tool for Transiting Exoplanets-report by excluding false positives, studying the background, satellite and centroid motion, and analyzing the pixels in sector 55 to define the source goodness. The candidate exoplanet transit has a duration of TD = 3.957 hr, Depth = 0.1196% and mid-transit time occurs at TBJD = 2811.24. Only one event was observed, so the orbital period cannot be evaluated.

Sustained attention and the flash grab effect.

When a stationary target is briefly presented on top of a moving background as it reverses direction, the target is displaced perceptually in the direction of the upcoming motion (the flash grab effect). To determine the role of attention in this effect, we investigated whether the predictability of the location of the flash grab target modulates the illusion. First, we established that effect was weaker for spatially predictable targets. Next, we showed that the flash grab effect decreased for a narrower spatial spread of attention before the onset of the target and that it was smaller for left hemifield presentations than right. Finally, we demonstrated that diverting attention away from the target and the background motion decreases the flash grab effect. In the first two experiments, the decrease in the illusion could be attributed to either increased attention to the target or decreased attention to the motion; we assume that increasing attention to the target necessarily decreases attention to the motion. However, in the final experiment, the central task decreases attention to both the target and the motion. The results show a decrease in the illusion and that reveals that attention to the motion is the primary causal factor.

The effect of background information and motion speed on the performance of TTC estimation

BackgroundIn previous studies, most research on motion perception have been conducted under background-free condition when the stimulus moved in a plane parallel to the observer. In real-life situations, people’s perception of the motion state of objects is usually done under different visual noise. Based on the occlusion paradigm, this study aimed to investigate whether different background information and motion speed affect the trend and accuracy of time-to-collision (TTC) estimation when stimuli move in a plane parallel to the observer.MethodsThirty five college students (mean age = 20.94, SD = 2.95, range = 18-28 years) participated in experiment 1, and used a 2 (background orientation: horizontal, vertical) × 3 (motion speed: slow, medium, fast) design to explore the effect of different line segment orientations and motion speed on TTC estimation performance; 36 college students (mean age = 20.81, SD = 2.82, range = 18-28 years) participated in experiment 2, and used a 2 (background dimension: two-dimensional background, three-dimensional background) × 3 (motion speed: slow, medium, fast) design to explore the effect of different background dimensions and motion speed on the performance of TTC estimation. The data were analyzed using SPSS 25.0.ResultsThe results revealed that: (1) The TTC was underestimated for the slow speed condition and overestimated for the medium and fast speed conditions. (2) The highest accuracy of TTC estimation was obtained for the fast condition. (3) The TTC were overestimated for the vertical background condition and underestimated for the horizontal background condition. (4) Compared to the two-dimensional background, the TTC was overestimated in the three-dimensional background.ConclusionsObject motion speed affected the TTC estimation performance, and different background information affected the TTC estimation performance when the object moved in a plane parallel to the observer. Meanwhile, the impact of background orientation and motion speed showed significant interactions.

An Angular Acceleration Based Looming Detector for Moving UAVs.

Visual perception equips unmanned aerial vehicles (UAVs) with increasingly comprehensive and instant environmental perception, rendering it a crucial technology in intelligent UAV obstacle avoidance. However, the rapid movements of UAVs cause significant changes in the field of view, affecting the algorithms' ability to extract the visual features of collisions accurately. As a result, algorithms suffer from a high rate of false alarms and a delay in warning time. During the study of visual field angle curves of different orders, it was found that the peak times of the curves of higher-order information on the angular size of looming objects are linearly related to the time to collision (TTC) and occur before collisions. This discovery implies that encoding higher-order information on the angular size could resolve the issue of response lag. Furthermore, the fact that the image of a looming object adjusts to meet several looming visual cues compared to the background interference implies that integrating various field-of-view characteristics will likely enhance the model's resistance to motion interference. Therefore, this paper presents a concise A-LGMD model for detecting looming objects. The model is based on image angular acceleration and addresses problems related to imprecise feature extraction and insufficient time series modeling to enhance the model's ability to rapidly and precisely detect looming objects during the rapid self-motion of UAVs. The model draws inspiration from the lobula giant movement detector (LGMD), which shows high sensitivity to acceleration information. In the proposed model, higher-order information on the angular size is abstracted by the network and fused with multiple visual field angle characteristics to promote the selective response to looming objects. Experiments carried out on synthetic and real-world datasets reveal that the model can efficiently detect the angular acceleration of an image, filter out insignificant background motion, and provide early warnings. These findings indicate that the model could have significant potential in embedded collision detection systems of micro or small UAVs.

Predicting cervical cancer target motion using a multivariate regression model to enable patient selection for adaptive external beam radiotherapy

Background and purposeInterfraction motion during cervical cancer radiotherapy is substantial in some patients, minimal in others. Non-adaptive plans may miss the target and/or unnecessarily irradiate normal tissue. Adaptive radiotherapy leads to superior dose-volume metrics but is resource-intensive. The aim of this study was to predict target motion, enabling patient selection and efficient resource allocation. Materials and methodsForty cervical cancer patients had CT with full-bladder (CT-FB) and empty-bladder (CT-EB) at planning, and daily cone-beam CTs (CBCTs). The low-risk clinical target volume (CTVLR) was contoured. Mean coverage of the daily CTVLR by the CT-FB CTVLR was calculated for each patient. Eighty-three investigated variables included measures of organ geometry, patient, tumour and treatment characteristics. Models were trained on 29 patients (171 fractions). The Two-CT multivariate model could use all available data. The Single-CT multivariate model excluded data from the CT-EB. A univariate model was trained using the distance moved by the uterine fundus tip between CTs, the only method of patient selection found in published cervix plan-of-the-day studies. Models were tested on 11 patients (68 fractions). Accuracy in predicting mean coverage was reported as mean absolute error (MAE), mean squared error (MSE) and R2. ResultsThe Two-CT model was based upon rectal volume, dice similarity coefficient between CT-FB and CT-EB CTVLR, and uterine thickness. The Single-CT model was based upon rectal volume, uterine thickness and tumour size. Both performed better than the univariate model in predicting mean coverage (MAE 7 %, 7 % and 8 %; MSE 82 %2, 65 %2, 110 %2; R2 0.2, 0.4, −0.1). ConclusionUterocervix motion is complex and multifactorial. We present two multivariate models which predicted motion with reasonable accuracy using pre-treatment information, and outperformed the only published method.

Bio-Inspired Small Target Motion Detection with Spatio-Temporal Feedback in Natural Scenes.

Small moving objects at far distance always occupy only one or a few pixels in image and exhibit extremely limited visual features, which bring great challenges to motion detection. Highly evolved visual systems endow flying insects with remarkable ability to pursue tiny mates and prey, providing a good template to develop image processing method for small target motion detection. The insects' excellent sensitivity to small moving objects is believed to come from a class of specific neurons called small target motion detectors (STMDs). However, existing STMD-based methods often experience performance degradation when coping with complex natural scenes. In this paper, we propose a bio-inspired visual system with spatio-temporal feedback mechanism (called Spatio-Temporal Feedback STMD) to suppress false positive background movement while enhancing system responses to small targets. Specifically, the proposed visual system is composed of two complementary subnetworks and a feedback loop. The first subnetwork is designed to extract spatial and temporal movement patterns of cluttered background by neuronal ensemble coding. The second subnetwork is developed to capture small target motion information where its output and signal from the first subnetwork are integrated together via the feedback loop to filter out background false positives in a recurrent manner. Experimental results demonstrate that the proposed spatio-temporal feedback visual system is more competitive than existing methods in discriminating small moving targets from complex natural environments.

Target motion misjudgments reflect a misperception of the background; revealed using continuous psychophysics.

Determining the velocities of target objects as we navigate complex environments is made more difficult by the fact that our own motion adds systematic motion signals to the visual scene. The flow-parsing hypothesis asserts that the background motion is subtracted from visual scenes in such cases as a way for the visual system to determine target motions relative to the scene. Here, we address the question of why backgrounds are only partially subtracted in lab settings. At the same time, we probe a much-neglected aspect of scene perception in flow-parsing studies, that is, the perception of the background itself. Here, we present results from three experienced psychophysical participants and one inexperienced participant who took part in three continuous psychophysics experiments. We show that, when the background optic flow pattern is composed of local elements whose motions are congruent with the global optic flow pattern, the incompleteness of the background subtraction can be entirely accounted for by a misperception of the background. When the local velocities comprising the background are randomly dispersed around the average global velocity, an additional factor is needed to explain the subtraction incompleteness. We show that a model where background perception is a result of the brain attempting to infer scene motion due to self-motion can account for these results.

Simulation of soft tissue deformation under physiological motion based on complementary dynamic method

Background and ObjectivePhysiological motions have a significant impact on soft tissue deformation and accuracy of surgical procedures, which is essential for realistic surgical simulation. While existing studies offer accurate simulation of soft tissue deformation, integrating physiological motions into deformation models of soft tissue remains a challenging task. MethodsThis paper introduces a novel deformation model, based on complementary dynamics, to animate soft tissue deformation under physiological motion. The finite element method is incorporated to accurately characterize the elastic behavior of the soft tissue. Mathematical models of physiological motion are utilized and the physiological effects are converted into displacements of a predefined set of handles within the soft tissue mesh. Complementary displacements derived from the inherent dynamics of the soft tissue are calculated, enabling the simulation of physiological motions and elastic behaviors in soft tissue deformation. ResultsExperiments were conducted to evaluate the performance and effectiveness of the proposed method in simulating soft tissue deformation under physiological motion. The simulation results show that the soft tissues exhibit physiological motion that corresponds to the rhythm of arterial pressure fluctuations, heartbeat or respiratory. Furthermore, the presented method exhibits stable performance compared with existing force-based methods. ConclusionsBoth elastic behaviors and physiological motions of soft tissue deformation can be governed by the proposed method. A high degree of realistic visualization is achieved for virtual surgery simulation.

A Technical Analysis of Digital Image and Video Processing

Video processing is the most emergent area of image processing in Research and Development. This paper, we present some advanced technology interrelated to video and image processing, which will provide a super-resolution high-quality most prominent visible light video for better human visualization and background motions or gesticulations of the human brain. There are two approaches for increasing the current resolution level of an image. The first one is the improvement of spatial resolution by reducing the size of pixels through the techniques of sensor manufacturing by about 40 M m 2 for a 0.35 M m CMOS process and the second approach is the enhancement of chip size that also enhances the capacitance. In video processing, we mainly focus on object images, i.e. any form of signal processing in which the input is an image and the output is a photograph or a frame of video. So, it will be better to converge over techniques of digital image processing. In the proposed algorithm (MHI Simulink model), edge extraction is performed to find the precious area of MRI and CT-Scan images. Using FPGA, edge detection methods have been implemented over scanned images to get better accuracy. The accuracy of the proposed method has been compared with Prewitt and Sobel’s edge detection techniques. The proposed method has given better accuracy than Prewitt and Sobel methods. Finally, this paper shows future directions for researchers to enhance the characteristics of digital image and video processing.

Co-occurrence spatial–temporal model for adaptive background initialization in high-dynamic complex scenes

Background information is an important aspect of pre-processing for advanced applications in computer vision. The literature has made rapid progress in background initialization. However, background initialization still suffers from high-dynamic complex scenes, such as illumination change, background motion, or camera jitter. Therefore, this study presents a novel Co-occurrence Spatial–Temporal (CoST) model for background initialization in high-dynamic complex scenes. CoST achieves a spatial–temporal model through a co-occurrence pixel-block structure. The proposed approach extracts the spatial–temporal information of pixels to self-adaptively generate the background without the influence of high-dynamic complex scenes. The efficiency of CoST is verified through experimental results compared with state-of-the-art algorithms. The source code of CoST is available online at: https://github.com/HelloMrDeng/CoST.git.

Novel simulation for dosimetry impact of diaphragm respiratory motion in four-dimensional volumetric modulated arc therapy for esophageal cancer

Background and purposeThe diaphragm respiratory motion (RM) could impact the target dose robustness in the lower esophageal cancer (EC). We aimed to develop a framework evaluating the impact of different RM patterns quantitatively in one patient, by creating virtual four-dimensional computed-tomography (v4DCT) images, which could lead to tailored treatment for the breathing pattern. We validated virtual 4D radiotherapy (v4DRT) along with exploring the acceptability of free-breathing volumetric modulated arc therapy (FB-VMAT). Methods and MaterialsWe assessed 10 patients with superficial EC through their real 4DCT (r4DCT) scans. v4DCT images were derived from the end-inhalation computed tomography (CT) image (reference CT) and the v4DRT dose was accumulated dose over all phases. r4DRT diaphragm shifts were applied with magnitudes derived from r4DCT scans; clinical target volume (CTV) dose of v4DRT was compared with that of r4DRT to validate v4DRT. CTV dosage modifications and planning organ at risk volume (PRV) margins of the spinal cord were examined with the diaphragm movement. The percentage dose differences (ΔDx) were determined between the v4DRT and the dose calculated on the reference CT image. ResultsThe CTV ΔDx between the r4DRT and v4DRT were within 1% in cases with RM ≦ 15 mm. The average ΔD100% and ΔDmean of the CTV ranging from 5 to 15 mm of diaphragm motion was 0.3% to 1.7% and 0.1% to 0.4%, respectively. All CTV index changes were within 3% and ΔD1cc and ΔD2cc of Cord PRV were within 1%. ConclusionWe postulate a novel method for evaluating the CTV robustness, comparable to the conventional r4DCT method under the diaphragm RM ≦ 15 mm permitting an impact of within 3% in FB-VMAT for EC on the CTV dose distribution.

Impact of Visually Induced Motion Sickness from VR Depending on Viewing Patterns, View Movement, and Background Motion.

The impact of visually induced motion sickness from virtual reality (VR) because of viewing patterns, view movements, and background global motion was investigated experimentally through classification into four categories.Each of the ten subjects underwent watching four patterns with bio-signal measurements, such as electrocardiogram and respiration, answering a subjective questionnaire.The results of the subjective evaluation of biological effects indicate that the VR viewing pattern has a significantly lower influence on visually induced motion sickness than other patterns, such as viewing with global background motion or sight motion by operating the controller.Clinical Relevance- This study aimed to reduce the effects of visually induced motion sickness in VR viewing, particularly to reduce the burden on the eyes and body.

Behavior automatic analysis for wolf pack hunting: making fast behavior analysis of massive data possible

Automatic wolf pack hunting behavior analysis contributes to the identification of behavioral mechanism and the development of bio-swarm intelligence engineering. However, wolf pack hunting commonly involves a complicated background and high-speed motion, where the frequent interactions with each other lead to frequent visual occlusion of the individual wolves. These difficulties make the automatic analysis of group hunting behavior significantly more challenging. Hence, we develop an automatic wolf pack hunting behavior detection scheme appropriate for videos, comprising a four-level individual feature map (frame, video, space, and semantic) and a group feature map. We propose a temporal–spatial feature fusion-based motion state recognition algorithm appropriate for scale-varied and occlusion-frequent individuals to obtain a robust semantic-level feature. Based on this individual feature map, we propose a video-based wolf pack hunting automatic behavior detection method. The developed scheme is validated on our Wolf2022 dataset, while the proposed motion state recognition and group behavior detection algorithms are further tested through ablation experiments. The results revealed that the motion state recognition accuracy reaches 88%, correctly detecting 15 out of 17 group behavior video clips.Significance statementIt is difficult to fast extract quantitative analysis results of wolf pack hunting behavior from video data. Our research focused on the research of the automatic analysis method for wolf pack hunting. Using the proposed method, the statical individual and group behavior attributes can be automatically generated from video data, which contributes to the building of a conceptual bridge between the wolf pack hunting behavior and bio-swarm intelligence engineering.