Hybrid Multiple Granularities Network and Hyperspherical Crown Probabilistic Neural Network with Class Aggregation for Person Re-Identification

With the increasing coverage of video surveillance equipment, person re-identification (Re-ID) is becoming more important in intelligent video surveillance system. This paper proposes the hybrid multiple granularities network (MGN) and hyperspherical crown probabilistic neural network (HCPNN) with class aggregation for person Re-ID. First, the discriminative global features and local features with various granularities extracted from pedestrian images extracted by MGN are concatenated to form the final feature representation. Then, the features are mapped with unit hyperspherical crown model and form the aggregation features for each class. HCPNN constructed with the aggregation features works as a classifier and we can get the proposed hybrid network. In the proposed method, nonlinear mapping model of HCPNN with class aggregation can improve the separability of the samples and minimize the impact of abnormal data as a classifier. Besides, we design the class aggregation strategy to reduce heavy computation of HCPNN under large samples effectively and maintain a high recognition rate. In addition, our experiments implemented on the mainstream datasets show that the proposed method have better performances than current person Re-ID methods.