Semantic Communication-Empowered Vehicle Count Prediction for Traffic Management

TL;DR

The paper addresses efficient real-time vehicle count prediction for smart-city traffic management by introducing a semantic communication framework. It presents a joint CNN-LSTM SemCom pipeline that extracts density-map semantics from raw images, transmits compact representations, and uses an LSTM-based decoder at the central controller to predict vehicle counts. Key contributions include a density-map semantic encoder, a partial residual connection, and end-to-end training with density and count losses; on TRANCOS, the approach achieves a 54.42% overhead reduction and outperforms state-of-the-art MAE/MSE benchmarks. This work demonstrates how semantic-by- design data reduction can enable scalable, low-latency traffic management in 6G-era networks.

Abstract

Vehicle count prediction is an important aspect of smart city traffic management. Most major roads are monitored by cameras with computing and transmitting capabilities. These cameras provide data to the central traffic controller (CTC), which is in charge of traffic control management. In this paper, we propose a joint CNN-LSTM-based semantic communication (SemCom) model in which the semantic encoder of a camera extracts the relevant semantics from raw images. The encoded semantics are then sent to the CTC by the transmitter in the form of symbols. The semantic decoder of the CTC predicts the vehicle count on each road based on the sequence of received symbols and develops a traffic management strategy accordingly. Using numerical results, we show that the proposed SemCom model reduces overhead by $54.42\%$ when compared to source encoder/decoder methods. Also, we demonstrate through simulations that the proposed model outperforms state-of-the-art models in terms of mean absolute error (MAE) and mean-squared error (MSE).

Figures (6)

• Figure 1: The sample traffic model of a smart city with entry ($A_i, i \in \{1, \ldots, 8\}$) and exit ($B_j, j \in \{1, \ldots, 8\}$) points. The main purpose of traffic management is to minimize the average travel time of a traveler to cross a city, i.e., the average time taken between an entry point $A_i, i \in \{1, \ldots, 8\}$ and an exit point $B_j, j \in \{1, \ldots, 8\}$.
• Figure 2: The block diagram of our proposed joint CNN-LSTM-based SemCom system model.
• Figure 3: The architecture of the semantic encoder, consisting of a CNN that produces the density map whose size is same as that of the input image.
• Figure 4: (a) Schematic diagram of a single LSTM cell. (b) The architecture of the semantic decoder, consisting of three layers of 100 LSTM cells.
• Figure 5: (a) This plot shows the loss function $\mathscr{L}_{count}$ versus the number of epochs for training and validation errors. (b) This plot shows the MAE values versus the hyper-parameter $p$. The minimum MAE value of $6.23$ is obtained when $p=0.8$.