Research on parking space detection algorithm in complex environments based on improved YOLOv7

Abstract In the context of complex parking environments, vehicle parking space detection faces challenges such as multi-scale, multi-angle, and occlusion issues, leading to low detection efficiency and problems with false positives and false negatives. In this study, we propose an improved vehicle parking space detection algorithm based on YOLOv7. Firstly, we enhance the convolutional layers by introducing the Mish activation function, thereby improving the model’s feature extraction capabilities and its ability to represent objects effectively. Secondly, we combine the parameter-free attention mechanism SimAM with feature pyramid modules and feature extraction modules to replace certain convolutional layers, thereby enhancing the reinforcement of critical parking space information and adaptability to variations in target scale. Finally, we replace the nearest-neighbor interpolation in the upsampling section with the lightweight operator CARAFE, effectively extracting parking space feature information and enhancing the algorithm’s feature fusion capabilities. Through ablation experiments and comparative trials on publicly available parking space datasets, our improved YOLOv7 algorithm achieves an mAP of 78.7%. Compared to the original algorithm, it demonstrates a 1.5% improvement in detection performance and a 5.6% increase in recall rate. These enhancements significantly improve parking space detection in complex environments, addressing issues such as false positives and false negatives, thereby meeting the performance requirements of parking space detection in parking lots.