PointDet: An Object Detection Framework Based On Human local Features In The Task Of Identifying Violations

Abstract

Object detection algorithms play an important role in the field of violation detection. However, small target detection is full of challenges in scenes related to human. The relationship between objects is usually not considered in object detection algorithms, which will make the model over relay on the high-order features and not make full use of local features. To address this issue, a novel framework named PointDet is proposed to learn local features which optimizes the detection effect of small targets in chemical plants. Since most of the targets to be detected in our dataset are highly correlated with human, human local features are used when designing the framework. First, we use a trained pose estimation model to extract local key point features. However, if local features are used directly, the relationship between them cannot be fully considered. Based on this situation, we have designed the one-vs-others module and the adaptive-graph-convolution module to reconstruct local features. In addition, for the output layer, the most challenging problem is how to better detect small targets. In our task, various small objects such as gloves, goggles, etc. have an obvious positional relationship with the local features of human body. In the output layer, we have designed a head attention module to make full use of this situation to optimize the small target detection problem. Specifically, our framework significantly outperforms state-of-the-art by 7.8 AP scores on field work dataset in chemical plants.