International Journal of Sustainable Building Technology and Urban Development ISSN:2093-761X(Print) 2093-7628(Online)

One of the major challenges of smart-city surveillance is the ability to detect the pedestrian behavior in large and congested cities. Our system is a hybrid deep-learning system, based on Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. This mixture detects spatial and time trends of pedestrian videos. The Grasshopper Optimization Algorithm is utilized to optimize hyperparameters and choose the most suitable features that help to approach convergence better and increase the precision of predictions. Explainable AI methods enhance the transparency of the safety-critical urban environment. We use Grad-CAM to raise spatial areas and SHAP to measure the contribution of each feature into predictions. The framework was benchmarked on JAAD, PETS, CASIA and CUHK Avenue. The accuracy of the results is 96.8 percent and the F1 score is 0.961 and the generalization is strong with the datasets. Our method is healthier when compared to the conventional CNNLSTM models, and other deep-learning baselines, and has less false-positive, as well as being more interpretable. This solution is one of the effective and clean real-time solutions to pedestrian monitoring, which ensures developing a safer and more sustainable movement in urban environments.