Research on Real-time Detection of Stacked Objects Based on Deep Learning

Deep Learning has garnered significant attention in the field of object detection and is widely used in both industry and everyday life. The objective of this study is to investigate the applicability and targeted improvements of Deep Learning-based object detection in complex stacked environments. We analyzed the limitations in practical applications under such conditions, pinpointed the specific problems, and proposed corresponding improvement strategies. First, the study provided an overview of recent advancements in mainstream one-stage object detection algorithms, which included Anchor-based, Anchor-free, and Transformer-based architectures. The high real-time performance of these algorithms holds particular significance in practical engineering applications. It then looked at relevant technologies in three emerging research areas: Parts Recognition, Intelligent Driving, and Agricultural Picking. The study summarized existing limitations in real-time object detection within complex stacked environments and provided a comprehensive analysis of prevalent improvement strategies such as multi-level feature fusion, knowledge distillation, and hyperparameter optimization. Finally, after analyzing the performance of recent advanced one-stage algorithms on official datasets, this paper conducted empirical tests on a self-constructed industrial stacked dataset with algorithms of different structure and analyzed the experimental results in detail. A comprehensive analysis shows that Deep Learning-based object detection algorithms offer extensive applicability in complex stacked environments. In addressing diverse target sizes, overlapping occlusions, real-time constraints, and the need for lightweight solutions in complex stacked environments, each improvement strategy has its own advantages and limitations. Selecting and integrating appropriate enhancement strategies is critical and typically requires holistic evaluation, tailored to specific application contexts and challenges.