Fast TLAM: High-precision Fine Grain Smoking Behavior Detection

This paper proposes a fast two-level attention model (Fast TLAM) smoking behavior detection network to detect smoking behavior. The Fast TLAM network is mainly divided into the following three phases: 1. Pre-processing stage: EdgeBox candidate region selection algorithm is used to generate a large number of candidate regions and then filter them;, candidate regions containing foreground objects will be reserved for transmission to object-level and local-level models;2. Object-level model: a CNN network is trained to filter and classify candidate regions in the preprocessing stage; the network is also trained to filter out background information, leave only patches containing the target to be detected, and abtain classification results; 3. Local level model: (1) a network is trained to classify candidate regions in the preprocessing stage; (2) candidate regions screened at the object level are clustered with K-means algorithm and then classified. Finally, the classification results are obtained. The classification results of the first and second stages are categorized to complete the entire detection process. Tests are carried out on a self-made experimental data set. Experimental results show that the Fast TLAM network has a very high accuracy rate of 92.68% can be identified only by object-level graphics, and does not need labeling information at all. Moreover, the network solves several defects, namely, low accuracy, high cost, and poor convenience of the traditional smoking behavior detection method.