Roadside Distributed Acoustic Sensing (DAS), being capable of capturing car-induced subsurface strain at high spatial-temporal resolution, has shown potential to revolutionize urban traffic monitoring by providing instantaneous updates on traffic flow, and possibly traffic accidents to optimize traffic management. However, due to noise contamination and interference among closely traveling cars, car detection and tracking accuracy is limited. To address these challenges, we design and train a self-supervised U-Net model that compresses car-induced DAS signals into sharp pulses through a spatial deconvolution. The localized and narrow outputs from our model lead to accurate and highly resolved car position and speed tracking, which can help identify driving behaviors and detect sudden stops due to accidents. We evaluate the effectiveness and robustness of our method through field recordings under different traffic conditions and various driving speeds. Our results show that our method can enhance the spatial-temporal resolution and better resolve closely traveling cars, which is beneficial for detecting and tracking cars under heavy traffic conditions and enables the characterization of large-size vehicles such as retrieving the number of bus axles and train bogies.