Högskolan i Skövde

Urban roundabouts present complex interaction scenarios where Advanced Driver Assistance Systems (ADAS) often generate excessive false alarms, undermining driver trust and system effectiveness. This master’s thesis investigates whether integrating spatiotemporal context into a deep sequential model can improve collision-risk estimation and reduce unnecessary warnings in cyclist–vehicle encounters. Building on the hybrid Decision Tree + Random Forest approach of Atif et al. (2025), we implement a Long Short-Term Memory (LSTM) network whose architecture is fine-tuned via Neural Architecture Search (NAS). The LSTM ingests enriched trajectory data—incorporating traffic density, proximity cues, and lane alignment— from stereo-camera recordings at Gothenburg roundabouts (nearly two million observations). Performance is evaluated against the baseline using five-fold cross-validation, focusing on accuracy, recall, precision, and false-alarm rate. Results show the NAS-optimized LSTM achieves 95% accuracy and reduces false positives by 46% compared to the Random Forest baseline, while maintaining balanced recall (0.93). To ensure transparency, we apply SHAP and LIME explainability methods, revealing that collision predictions emerge from subtle temporal patterns rather than isolated static features. While the Random Forest excels in spatial interpretability, the LSTM’s sequence-level reasoning drives its superior false-alarm reduction. We discuss methodological trade-offs, generalizability to diverse urban environments, and ethical considerations—such as data privacy under GDPR and the risk of driver over-reliance on automation. These findings suggest that context-aware deep learning, when paired with explainable AI, can meaningfully enhance ADAS reliability and support safer, more trustworthy urban mobility.