Blind-Aided Target Detection Algorithm Based on Cascading Feature Pyramids With Lightweight Dual-Path Downsampling

Owing to the high frequency of missed detection of small targets in complex environments, insufficient multi-scale target feature extraction, and high computational complexity of the algorithm, it is a difficult challenge for intelligent blind guide devices in the field of target detection algorithms. To solve the existing problems and challenges, this study proposes a Kinective-Journey YOLO (KJ-YOLO) target detection algorithm customized for target detection in complex environments. It uses the idea of cascaded feature pyramids to enhance the expression ability of feature maps through an attention mechanism and feature weighting to realize a more complex feature fusion mechanism: the Scale Balancing Aggregation Module (SBA Module). It improves C3k2 through an intelligent segmentation and reorganization mechanism of feature channels and residual cross-layer connections to achieve feature reuse, thereby reducing computational complexity and improving the C3K2-DEConv module of small-target feature extraction. The feature extraction and fusion of small and multi-scale targets are improved, and the lightweight downsampling module Adown is introduced to reduce the computational complexity of the model. A large number of experiments on the PASCAL VOC 2007 public dataset show that, compared with the baseline model, KJ-YOLO’s ‘mAP’ index rate is increased by 3.3% to 67.3%, and mAP50-95 is increased by 2.3% to 44.8%. The number of parameters and the amount of calculation are also controlled at 2.9M and 114.4 FLOPs/G, which not only improves the accuracy but also meets the requirements of lightweight deployment and real-time detection of a smart device for blindness assistance.