| Pedestrian detection for vehicle-mounted surveillance video means that the camera device installed on the driving vehicle is applied to obtain the current road video information firstly.The computer vision technology is further adopted to detect and locate the spatial position of pedestrians from the video sequence.With the development of driverless technology,driverless cars have appeared in the real society.And as a vulnerable group in the traffic environment,the safety guarantee for pedestrians cannot be ignored.Space scale variation of pedestrian instance is one of the main bottlenecks affecting pedestrian detection performance.Based on the feature differences among different scale spaces,a scale awareness strategy based on multi-path Region Proposal Network(RPN)is firstly introduced.And according to the effectiveness of the convolutional feature layers with different scales,a series of candidate regional scale sets are generated adaptively from the feature maps corresponding to the size of the receptive field.Secondly,considering the semantic complementarity of convolutional features at different levels,a trans-scale feature aggregation network(TS-FAN)is proposed to enhance representation ability of convolutional features by aggregating horizontal connection,top-down path and bottom-up path.Finally,the multi-path RPN scale awareness strategy and trans-scale feature aggregation model are combined to construct a multi-scale pedestrian detection network by adaptive scale perception.Experimental results show that the log-average miss rate of the proposed method on Caltech benchmark has reduced to 5.53% for scale pedestrian(more than 50 pixels in height),and 26.21% for whole scale pedestrians(more than 20 pixels in height).And The proposed metric has got a better performance on ETH benchmark with drastic scale variations.A large number of low-resolution pedestrian vehicle-mounted surveillance videos is one of the main bottlenecks affecting pedestrian detection performance.Therefore,a low-resolution pedestrian detection network based on self-attention mechanism is designed in this paper to effectively improve the detection effect of low-resolution pedestrians.Firstly,the trans-scale feature aggregation network module is developed to make full use of the aggregation features with strong characterization ability and high robustness by discarding the original top-down connection path,rather than integrating the aggregation features with low-level resolution features.Then,in order to retain the local correlation between channels,group self-attention mechanism is employed to recalibrate the feature layers used for classification and regression.And to explore the potential area of pedestrians and suppress the interference of background information,average pooling and maximum pooling are used to obtain the attention representation vector to make up for the deficiency of the single form.Experimental results show that the average miss rate of the proposed network has reduced to 22.85% on the All testing dataset in the Caltech benchmark,and to 41.99% for small scale pedestrians(pedestrian height between 20 and 30 pixels).Compared with state-of-the-art pedestrian detection methods on the datasets of pedestrians with different resolutions,the proposed method has improves the pedestrian detection effect of low-resolution in driving surveillance video and has good robustness and high recall rate. |
