DATA ASSOCIATION USING CORRELATION RESPONSE VALUES

Smart overview of the Invention

In the realm of visual perception, efficient multi-object tracking is crucial, especially in applications like surveillance, animation, and vehicle navigation. Traditional approaches often rely on single-object trackers processing video frames independently, which can be computationally expensive and inefficient in crowded or occluded environments. The disclosed system aims to enhance tracking by utilizing correlation filters, enabling more effective localization and data association across frames.

Parallel Processing

The disclosed method extracts image areas from a batch of images and scales them to template sizes for efficient processing. By loading these scaled search regions into GPU memory, they can be processed in parallel, significantly boosting localization and filter update efficiency. This parallel processing capability allows for handling data from multiple images and trackers simultaneously, optimizing computational and storage resources.

Correlation Response Values

The system introduces a novel approach for associating object locations using correlation response values. By determining an estimated location of an object through correlation filters, the system uses these response values as visual features, eliminating the need for separate visual feature extraction. This streamlines the process of linking object locations across frames, enhancing tracking accuracy and efficiency.

Focused Windowing and Occlusion Maps

To improve the learning of correlation filters, the system employs focused windowing and occlusion maps. Focused windowing applies a Gaussian filter to blur the background around a target object, reducing background influence while maintaining context. Occlusion maps mask or blur occlusions, allowing the correlation filter to focus on unobstructed parts of the target object. These techniques ensure robust learning and tracking, even in complex environments.

Applications and Implementation

The disclosed methods are applicable across various perception-based systems, including automotive, robotics, and smart area monitoring. By making image areas more homogenous and leveraging GPU parallelization, the system enhances processing efficiency. These improvements are adaptable to both autonomous and semi-autonomous vehicles, as well as other machine systems, providing a versatile solution for modern object tracking challenges.