Accurate forecasting of short-term traffic is important for traffic management and planning in intelligent transportation systems (ITS). We introduce a computer vision-guided deep learning pipeline using a hybrid and multi-layered architecture to extract intricate spatiotemporal features from city traffic data. It uses Convolutional Neural Networks (CNNs) to extract spatial characteristics and subsequently utilizes Shuffle Attention (SA) to enhance spatial information. The output of the SA module is input into an attention mechanism to enhance and accentuate the spatial elements. We employ a Long Short Term Memory (LSTM) network with an attention module. This module extracts and prioritizes spatial and short-term temporal features. We also incorporate a Bidirectional Long Short-Term Memory (Bi-LSTM) unit to enhance temporal feature extraction to capture long-term patterns. The method is improved by adding another attention mechanism to concentrate on details, like changes in traffic patterns between weekdays and weekends. The effectiveness of this approach has been confirmed using traffic video data gathered during 36 h. Experiment results validate that the proposed hybrid model outperforms existing methods with better accuracy, convergence speed, and adaptability for short-term traffic flow prediction, especially in complex weather scenarios. The code supporting this research is available at https://github.com/arabindaiitbbs/TrafficFlow_HybridModel_CVIP2024.git .