Shoplifting activities pose a significant challenge for retail stores and law enforcement agencies. Stakeholders lose billions of dollars to such activities every year. Surveillance cameras can detect such shoplifting activities, yet human limitations such as visual focus over prolonged periods and the rarity of such activities decrease the detection probability. Several researchers have proposed deep learning-based solutions to utilize surveillance videos for shoplifting detection. However, most have evaluated their approaches on traditional datasets generated under controlled environments. Furthermore, identifying shoplifting activities in real-time video sequences is still challenging, as the details in video sequences have a time continuity constraint. To this end, the current work proposes a two-stage shoplifting detection approach, where the first stage employs a video pre-processing technique referred to as “pixel-difference” that discards the extraneous objects from the video frames and the second stage employs a deep learning model “P3DConvMemoryNet” based on 3DCNN and ConvLSTM layer to learn on the spatiotemporal features for detection. Moreover, the current work uses real-time data collected from four retail stores in an uncontrolled environment to evaluate the proposed approach. The experimental results reveal that the proposed approach exhibits the F1-scores of 0.75, 0.71, 0.73 and 0.81 for the four stores, respectively. Additionally, the results validate the efficacy and applicability of the proposed approach in different store environments with minimal modifications to the architectural configurations.