Recently, there has been an increase in the so-called "dead drop" method, where a dropper leaves drugs at a predetermined location and the buyer picks them up later. As illegal drug trades like dead drops increase, the domestic drug problem is becoming more severe. To address this issue, we intend to use Atomic action recognition to record and extract suspicious dead drop behavior and the suspect's ID, then input this data into a sequential model-based Composite action recognition model. As the atomic actions of a specific suspect accumulate, the current composite action state is estimated based on this information to recognize dead drops.  

Atomic Action Recognition is a subfield of action recognition that focuses on identifying and classifying the smallest possible units of action or movement within a sequence. These atomic actions are the fundamental building blocks of more complex activities. For example, in the context of human activities, atomic actions might include movements like raising a hand, taking a step, or turning the head. In this project, we aim to detect dead drops by identifying atomic actions such as digging, looking around, and checking a phone through Atomic Action Recognition, and ultimately using Composite Action Recognition. 

In this research, composite action recognition is employed to detect drug dead dropping. Initially, information such as action labels, person IDs, and action durations (start time and end time) is extracted through previously performed atomic action recognition. This extracted information is then analyzed using sequential models such as LSTM, GNN, and Transformer. By doing so, we identify patterns of actions that occur in sequence, such as loitering, taking photos, digging, burying, and placing objects in secluded areas. This method allows for the detection of drug dead dropping by recognizing complex action sequences that are not easily identified through atomic action recognition alone.