Real-time Pedestrian Detection Algorithm of YOLOv10 under Complex Lighting and Occlusion Conditions
Hengzhen Zhang · 2024
In order to enhance the efficiency and accuracy of real-time pedestrian detection, meeting the demands of intelligent traffic monitoring and autonomous driving technology, we employ an advanced algorithm based on YOLOv10 for pedestrian detection. The algorithm integrates the core features of YOLOv10, including a dual-label assignment system, compact inverted block (CIB), large-core convolution, and partial self-attention (PSA) mechanisms, among other modules, aiming to strengthen the model's adaptability and robustness to complex scenarios. This paper collects a pedestrian dataset from various scenes on campus and uses a visual large model for automatic data annotation, which is then verified manually and used for experiments after expansion. The experimental results show that YOLOv10 performs excellently in the expanded version of the Caltech pedestrian detection dataset, achieving an average precision mean (mAP) accuracy of up to 93.4%, which significantly improves detection performance compared to YOLOv8. The specific data shows that the B variant of YOLOv10 reached an mAP@50 score of 93.3%, while also being superior to the previous model in terms of the number of parameters (20.4M) and floating-point operations per second (FLOPs), proving its efficiency and accuracy in real-time detection tasks. These results not only confirm the ability of the YOLOv10 algorithm to handle occlusions, lighting changes, and the diversity of pedestrian postures in special scenarios but also demonstrate its great potential and value for promotion in practical applications.