Target detection method based on event-time surface feature fusion network

Target detection is a core task in computer vision, but traditional methods based on RGB images often encounter challenges such as computational complexity and high resource consumption. In contrast, event cameras have garnered significant attention due to their advantages, including high dynamic range, real-time performance, low data redundancy, and fast sensing. This paper introduces a novel target detection method grounded in event-time surface feature fusion. The proposed neural network architecture integrates four key modules: (1) extraction of biologically-inspired time-surface features from event cloud data by leveraging temporal information within local spatial neighborhoods; (2) utilization of a self-attention mechanism within the event cloud feature extraction network to capture point features; (3) fusion of event time-surface features with point features; and (4) sampling and grouping techniques to generate voting centers and clusters, with a proposal module providing final object localization and recognition. Experimental results demonstrate that this event-time surface-based method effectively harnesses raw event data, significantly improving detection accuracy.