Causality Based Vehicle Re-Identification Method

The increase in the number of vehicles not only puts the traffic control system in hot water but also profoundly endangers public security.It is costly and timeconsuming for the police to manually determine the exact location of vehicles driven by drivers who break the law or traffic rules.Additionally,it is not always possible to locate vehicles by license plate recognition in complex environments.At this point,vehicle re-identification becomes the primary means of tracking target vehicles,as it could match the same vehicle between different cameras and plays an important role in modern traffic management systems.Nowadays,vehicle re-identification methods based on deep learning are becoming mainstream.However,most of these methods are in a data-driven correlation mode with poor stability and interpretability,which could lead to various problems.Firstly,the distinguishing features learned from a large amount of data are sub-optimal and could make it difficult to identify similar vehicles.Secondly,the model may be unable to distinguish major cues from biased cues in vehicle images.Finally,the supervision signal of the model may be weak and cannot effectively guide the model training.To solve the above problems,it is necessary to construct a more robust vehicle re-identification model based on causality,the contributions of the paper could be summarized into the following two points:(i)A vehicle re-identification model based on metric learning and causality was constructed.To better extract rich structural and detailed information within the global scope of the vehicle,the Transformer is chosen as the basic infrastructure.To avoid losing local information at patch boundaries,targeted improvements are made to the Transformer reducing the sliding window step size used for patch division.Using counterfactual theory based on causality to improve the network framework,various counterfactual attentions are designed to stimulate the model to learn more effective real attentions and reduce predictions based on incorrect attentions by maximizing the difference between them and real attention.Based on causal effects,a more robust supervised signal,the counterfactual attention loss function,is constructed.The final loss function consisted of identity loss,triplet loss,and counterfactual attention loss.The adaptive loss function weight algorithm is designed to automatically adjust the loss function weight,to further balance the training process and improve the performance of the model.(ii)A vehicle re-identification model based on average accuracy and causality is constructed.Using Transformer as the underlying architecture,the traditional metric loss function is replaced by the average precision loss function,specifically,the temperature-scaled sigmoid function is used to replace the indicator function in the average precision calculus to improve the model directly by optimizing the evaluation metrics,solving the problem that the metric loss function cannot be optimized based on ranking.On this basis,the improved sliding window chunking approach module,counterfactual attention,counterfactual attention loss,and adaptive loss function weights are added to the model to further improve its performance.The results of the ablation experiments on the Ve Ri-776 and Vehicle ID datasets confirm the effectiveness of the proposed modules in this paper.Comparative experiments with state-of-the-art methods show that the two causality-based vehicle re-identification models proposed in this paper have advanced performance,and the causality-based vehicle re-identification models proposed in this paper introduce very little computational cost during training and almost none during testing compared with the original Transformer model.