A behaviour detection algorithm integrating lightweight networks and feature recombination

Traditional behaviour detection methods often have problems such as low accuracy and slow processing speed, making it difficult to meet the practical application needs of industrial production scenarios. This study proposes a behaviour detection algorithm that integrates lightweight networks and feature recombination. First, we replace you only look once (YOLO) backbone with an enhanced MobileNetV3, reducing model complexity and accelerating inference. Second, we introduce content-aware reassembly of features, replacing conventional upsampling to improve precision. Further, switchable atrous convolution in the neck network enhances adaptability to multi-scale features, while vision transformer with deformable attention strengthens spatial modelling. Ablation experiments demonstrate the algorithm's effectiveness, achieving a 75.2% mAP, with gains of 2.8% and 4.8% in precision and recall, respectively. Compared to existing technologies, this method offers the advantages of fast speed and high accuracy, making it suitable for real-time detection scenarios, such as those in the petrochemical industry.