Automated Labelling of Traffic Police Data Based on Weak Supervision

In the context of autonomous driving, the effective recognition and understanding of traffic police gestures is crucial for safe driving on the road. Pedestrians’ body movements affect the driving system’s understanding of traffic police gestures; therefore, the localization and recognition of traffic police is key to the understanding of traffic police gesture. With the development of deep learning technology, a satisfactory traffic police localization and recognition model requires a huge amount of accurately labelled datasets, but there are no publicly available relevant datasets, and the cost of accurate labelling of datasets is extremely high, limiting the performance of the model. To address this problem, this paper proposes a High-Quality Proposals generation algorithm based on weak supervision (HQP). HQP improves the quality of region proposal generation by effectively enhancing the classification ability of a student model through guidance from a pre-trained teacher model. To compensate for the lack of localization information in image-level annotation data, HQP integrates gradient-weighted class activation maps to improve the model’s localization ability. HQP effectively reduces the workload of accurate data labelling and completes the automated labelling of data on traffic police localization and positioning by combining manual fine-tuning methods.