BR-102024017639-A2 - METHOD FOR PREDICTIVE ANALYSIS OF BEHAVIOR AND TRAJECTORY THROUGH TEMPORAL MODELING, OPTIMIZATION OF EMERGING PATTERNS AND DATA INTERPOLATION

Abstract

The present invention relates to a predictive analysis method for behavior and trajectory using a multi-algorithm system divided into processing modules which integrates computer vision techniques, convolutional and recurrent neural network models, time series modeling, and bio-inspired optimization methodology. The synergistic combination of these algorithms, where HOG and YOLO ensure precise and efficient object detection, cubic spline interpolation ensures trajectory continuity, RNNs model temporal dependencies to predict future behaviors, Prophet adjusts predictions based on temporal trends, and ACO optimizes predictive trajectories in real time, provides a complex and, moreover, adaptive model, operating continuously and optimally in trajectory prediction in complex environments, minimizing errors and adjusting analyses as conditions evolve.

Inventors

• CLAUDIO DE SOUZA ROCHA JUNIOR
• ADILSON VILARINHO TERRA

Assignees

• CLAUDIO DE SOUZA ROCHA JUNIOR

Dates

Publication Date

20260310

Application Date

20240827

Claims (4)

1. 1) METHOD FOR PREDICTIVE ANALYSIS OF BEHAVIOR AND TRAJECTORY THROUGH TEMPORAL MODELING, OPTIMIZATION OF EMERGING PATTERNS AND DATA INTERPOLATION, comprising an algorithm for data analysis using artificial intelligence applied to the tracking and prediction of trajectories and behaviors, characterized by a predictive analysis system of behavior and trajectory, divided into three modules, as shown in FIG. 1, being: a) Backend Algo 1 - responsible for the initial processing of videos; b) Backend Algo 2 - responsible for tracking and prediction using YOLO, HOG, data interpolation, Prophet and ACO; c) Frontend Interface - responsible for displaying the results in a graphical interface;
2. 2) METHOD, according to claim 1, characterized by combining a trajectory prediction method that combines Histograms of Oriented Gradients (HOG), YOLO, Recurrent Neural Networks (RNNs), cubic spline interpolation, Prophet and the Ant Colony Method (ACO), where HOG is applied for object detection in each video frame, extracting visual features based on the contours and textures of the objects, followed by the application of YOLO in conjunction with HOG to refine object detection, identifying the bounding boxes of the objects in real time and applying Non-Max Suppression to eliminate redundant detections; After object detection, cubic spline interpolation is used to fill gaps in the objects' trajectories, smoothing the data and ensuring that the trajectories are continuous and represent the movements of the tracked objects. These trajectories are then processed by Recurrent Neural Networks (RNNs), especially in the LSTM and GRU variations, which model the temporal dependencies in the trajectory data and make predictions about the future behavior of the objects. Prophet is integrated into the system to perform a time series analysis on the predictive data, adjusting the predictions based on long-term trends and seasonality. Finally, the Ant Colony Method (ACO) is applied to optimize predicted trajectories, using an iterative process that adjusts predictions based on emerging patterns observed in the trajectories, ensuring that predictions are adaptive in real time. This method is capable of operating continuously, with each component of the multi-algorithm interacting to detect, track, predict, and optimize object trajectories in complex environments, minimizing prediction errors and adjusting analysis strategies as environmental conditions evolve.
3. 3) METHOD, according to claim 1, characterized by using a real-time processing module with configurable latency, allowing the system to operate in both real-time and adjustable time intervals, ensuring flexibility in predictive analysis;
4. 4) METHOD, according to claim 1, characterized by using data interpolation techniques, including cubic spline interpolation, to ensure the continuity of the trajectories of the tracked objects, filling gaps in the data and smoothing the trajectories, in order to minimize the prediction error in dynamic and complex environments.

Description

FIELD OF THE INVENTION [001] The invention focuses on the technological field of data analysis and artificial intelligence applied to tracking and predicting trajectories and behaviors. In particular, the method integrates Histograms of Oriented Gradients (HOG) for visual detection, data interpolation for creating continuous trajectories, Recurrent Neural Networks (RNNs) for temporal modeling, and Ant Colony Methods (ACO) for optimizing emerging patterns. The main application is in environments such as retail, distribution, industry, and vehicle traffic, where accurate prediction and pattern analysis are essential for improving layout management and route optimization. TECHNICAL BASIS [002] Predictive analysis of behaviors and trajectories has become an area of great interest in various sectors, such as retail, distribution, industry, logistics, security, and transportation. With the advancement of data capture and processing technologies, there is a growing demand for solutions that can not only track objects in real time but also predict their future movements with high precision. To meet this demand, the combination of different techniques of computer vision, machine learning, and pattern optimization has proven to be a promising and effective approach. [003] One of the main techniques employed in object detection is the use of Histograms of Oriented Gradients (HOG), which has stood out for its ability to extract detailed visual features, such as contours and textures, from images. Introduced by Dalal and Triggs in 2005, HOG is widely used in object detection tasks, especially due to its effectiveness under varying lighting and perspective conditions. The HOG technique divides the image into blocks and calculates the histogram of gradient orientations in each block, thus capturing local features that are robust to the aforementioned variations. When combined with machine learning methods, such as convolutional neural networks (CNNs), this technique can be used to train models that identify and classify objects in real time, as in the case of the YOLO (You Only Look Once) algorithm. [004] YOLO is an approach that has revolutionized object detection by being extremely fast and efficient. Unlike previous methods, which divide the detection task into several steps, YOLO performs this task in a single pass through the neural network, allowing for real-time object detection. YOLO's speed is particularly useful in applications where response time is critical, such as in surveillance systems and in analyzing customer behavior in retail environments. The latest version of YOLO is capable of detecting multiple objects in an image, adjusting the bounding boxes and object classes with great precision, making it ideal for use in conjunction with other prediction and optimization techniques, such as those described in this invention. [005] Data interpolation plays a crucial role in the continuity of predictive analytics, especially in scenarios where captured data may have temporal gaps. Interpolation techniques, such as linear interpolation and cubic spline interpolation, allow these gaps to be filled, ensuring that object trajectories are continuous and accurate. This is particularly important in applications where trajectory accuracy is essential for decision-making, such as predicting customer flows in retail stores or modeling vehicle traffic in large cities. Spline interpolation, for example, is widely used due to its ability to smooth trajectories, minimizing prediction error and providing a more faithful representation of real movements. [006] To predict the future behavior of objects, Recurrent Neural Networks (RNNs) are a common and effective choice. RNNs are designed to process temporal data sequences, capturing the temporal dependencies between events. Among the most effective variations of RNNs are Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs), which are capable of storing information over long periods, overcoming the limitations of traditional RNNs, such as the vanishing gradient problem. In the context of this invention, RNNs are used to analyze continuous trajectories generated by data interpolation and predict the future movements of tracked objects, allowing for precise and detailed predictive analysis. [007] The Ant Colony Method (ACO) is an optimization technique inspired by the collective behavior of ants in nature, where ants find the shortest path between a food source and the nest using pheromones. In the context of predictive trajectory analysis, ACO is used to identify and optimize emergent patterns, adjusting predicted trajectories based on observed movement patterns. The application of ACO allows the system not only to predict the future behavior of objects, but also to optimize these predictions in real time, dynamically adjusting to changes in the environment. This adaptability is crucial in environments where movement patterns can be highly dynamic and unpredictable, suc