A Lightweight Temporal Detection Framework for Illegal Waste Dumping in Real Surveillance Footage

Illegal waste dumping poses significant environmental and economic challenges, and automated video-based detection systems have emerged as promising tools for supporting sustainable waste management. This paper presents a lightweight temporal deep learning framework for detecting waste dumping actions in surveillance videos, developed within the context of the Illegal Waste Dumping Challenge. Our approach combines convolutional feature extraction using a pretrained ResNet18 backbone with a bidirectional GRU-based temporal head. A dedicated dataset loader aggregates uniformly sampled frames from each video and generates frame-level supervision centred around the annotated dumping event. To improve robustness under strong class imbalance and noisy labels, we explore multiple labelling strategies, including single-frame labels, temporal window labels, and Gaussian soft labelling. Inference is performed by uniformly sampling a fixed-length clip from each video, ensuring consistency with the training procedure and enabling fast execution. While the current implementation avoids sliding-window inference to reduce latency and remain computationally efficient, it also supports deployment in constrained environments. The proposed framework naturally extends to sliding-window processing for longer videos, where multiple temporal segments can be analysed independently. Experimental results show that the system reliably identifies dumping events under temporal uncertainty, achieving competitive performance on a balanced validation set and demonstrating strong generalisation capabilities. This work provides insight into lightweight temporal modelling strategies for action spotting in real-world environmental monitoring scenarios.