Multi-Task Pedestrian Detection Network for Resource-Constrained Edge Devices

The patent merges the pedestrian location detection network and attribute recognition network into a unified multi-task recognition framework. Both tasks share a common backbone network for feature extraction, and the networks are trained jointly using a combined loss function. This integration allows the system to perform both location detection and attribute recognition simultaneously using a single network model, thereby reducing device resource consumption while maintaining detection accuracy.

The patent designs a multi-task recognition network that performs multiple functions simultaneously: pedestrian location detection, pedestrian attribute recognition, and auxiliary tasks such as pedestrian re-identification. The network processes a single input image to generate multiple output results, making the system more versatile and efficient by eliminating the need for separate processing pipelines for different detection tasks.

The patent combines multiple detection tasks into a single unified network architecture that processes images once and simultaneously outputs pedestrian locations, attributes, and other information. This merging eliminates the need for sequential processing through separate models, thereby improving detection efficiency and processing speed while maintaining task specialization through dedicated output layers for each task.

The patent implements continuous feature extraction and processing within a single network forward pass. The backbone network continuously extracts features that are then utilized by multiple task-specific heads simultaneously, ensuring that the useful action of feature processing continues without interruption or redundant computation, thereby maximizing detection efficiency.

The patent merges multiple detection capabilities into a single multi-task recognition network that handles pedestrian location detection, attribute recognition, and re-identification simultaneously. This consolidation reduces system complexity by eliminating the need to manage multiple separate models, their interfaces, and data flows, while maintaining comprehensive detection capability through integrated task processing.

The patent creates a universal recognition network that performs multiple detection functions through a unified architecture. The network uses shared feature extraction layers and task-specific output layers to achieve comprehensive detection capability, thereby simplifying the system structure compared to using separate specialized models for each detection task.