A Study on Suspect Identification in Videos Using Shoeprints

May 31, 2023

The utilization of shoe images for matching and locating suspects in video footage has become an important research area within the field of shoeprint analysis. The challenge lies in comparing high-resolution shoe images with low-resolution ones. To address this issue, we have developeda cross-resolution shoe retrieval algorithm based on deep learning. The aim is to enable comparison and retrieval between high-resolution shoe images and their low-resolution counterparts, thereby establishing a pathway from crime scene shoe print to shoe models, and eventually to video analysis.

      After years of refinement, our shoe model query product based on crime scene shoeprints has achieved a high level of maturity. This research focuses specifically on locating suspects in videos through the application of the shoe images comparison algorithm. During the algorithm implementation, we have introduced network branches, incorporated specific loss functions, and implemented random transformations.

      A total of 1,172 sample retrieval tests were conducted, with the results indicating a high level of effectiveness in both objective metrics and subjective verification. The test data consists of two parts. The first part comprises high-definition shoe model images captured using mobile phones, cameras, professional shoe model acquisition equipment, and selected online data. The second part consists of low-definition video shoe model images obtained from 4-megapixel camera footage of the shoes. Please refer to the figure below for more details.

Figure 1 Shoe model images
Table 1 Volume of test data

By leveraging a shoe recognition network model built on deep learning techniques, we extract distinctive features from both high-definition and low-definition shoe images.These extracted features serve as the basis for our subsequent calculations,comparisons, and ranking processes using a feature comparison algorithm. Through this methodology, we generate comprehensive retrieval results that enable effective matching and identification of shoe patterns.

        The combination of the shoe recognition network model and the feature comparison algorithm empowers our system to handle images of varying resolutions and extract meaningful information from them. By accurately comparing the extracted features, we can successfully retrieve relevant shoe patterns and provide valuable insights for locating suspects in videos. This approach show cases the potential of deep learning in addressing the challenges associated with shoe recognition and advancing thefield of shoeprint analysis.

Figure 2 Flowchart of Shoe image comparison and detection

In the retrieval test of 1,172 samples, the matching results were sorted based on the degree of feature matching, with the highest-ranking matches appearing first. Among the matching results, approximately 66.2% of the HD shoe samples were successfully matched with the first video shoe sample picture. Moreover, 85.8% of the HD shoe samples were successfully matched among the top 5 video shoe samples. Please refer to the figure below for a partial presentation of the test results.

Figure 3 Test result 1
Figure 4 Test result 2

Currently, our algorithm has reached a satisfactory level of application, and we continue to make improvements. We are enhancing the algorithm's compatibility with varying picture sharpness and shooting angles, expanding the dataset, and upgrading the deep learning network model.

        In addition to optimizing the algorithm, we have developed products based on this algorithm to serve users. Our products have received recognition and praise from a large number of users.The successful development of the algorithm for retrieving target individuals through shoes from extensive video data has significantly saved manpower and resources in the field of criminal investigation. It has also opened up new possibilities for the application of shoeprint analysis.




Cross-definitionShoe Upper Retrieval Based on Deep Learning – Implement the Application forLocking Suspects by Shoeprints in the Video.

Jin Yifeng, SunXirui, Wu Wenda, Li Daixi, Jiang Xuemei, Geng Xiaopeng.